Productivity underpins business success and national well-being and thus it is crucial to understand the factors that in

*431*
*92*
*17MB*

*English*
*Pages 856
[857]*
*Year 2018*

- Author / Uploaded
- Emili Grifell-Tatjé (editor)
- C.A. Knox Lovell (editor)
- Robin C. Sickles (editor)

T h e Ox f o r d H a n d b o o k o f

P RODU C T I V I T Y A NA LYSI S

The Oxford Handbook of

PRODUCTIVITY ANALYSIS Edited by

EMILI GRIFELL-TATJÉ, C. A. KNOX LOVELL, and

ROBIN C. SICKLES

1

3 Oxford University Press is a department of the University of Oxford. It furthers the University’s objective of excellence in research, scholarship, and education by publishing worldwide. Oxford is a registered trade mark of Oxford University Press in the UK and certain other countries. Published in the United States of America by Oxford University Press 198 Madison Avenue, New York, NY 10016, United States of America. © Oxford University Press 2018 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without the prior permission in writing of Oxford University Press, or as expressly permitted by law, by license, or under terms agreed with the appropriate reproduction rights organization. Inquiries concerning reproduction outside the scope of the above should be sent to the Rights Department, Oxford University Press, at the address above. You must not circulate this work in any other form and you must impose this same condition on any acquirer. Library of Congress Cataloging-in-Publication Data Names: Grifell-Tatjé, E. (Emili), editor. | Lovell, C. A. Knox, editor. | Sickles, Robin, editor. Title: The Oxford handbook of productivity analysis / edited by Emili Grifell-Tatjé, C.A. Knox Lovell, and Robin C. Sickles. Description: New York, NY : Oxford University Press, [2018] | Includes index. Identifiers: LCCN 2017049559 | ISBN 9780190226718 (hardcover : alk. paper) | ISBN 9780190226732 (epub) Subjects: LCSH: Industrial productivity. | Industrial productivity—Measurement. Classification: LCC HD56 .O95 2018 | DDC 658.5—dc23 LC record available at https://lccn.loc.gov/2017049559 1 3 5 7 9 8 6 4 2 Printed by Sheridan Books, Inc., United States of America

Contents

Contributors

ix

PA RT I E DI TOR S’ I N T RODU C T ION 1. Overview of Productivity Analysis: History, Issues, and Perspectives Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles

3

PA RT I I T H E F O U N DAT ION S OF P ROD U C T I V I T Y A NA LYSI S 2. Empirical Productivity Indices and Indicators Bert M. Balk

77

3. The US Bureau of Labor Statistics Productivity Program Lucy P. Eldridge, Chris Sparks, and Jay Stewart

121

4. Theoretical Productivity Indices R. Robert Russell

153

5. Dynamic Efficiency and Productivity Rolf Färe, Shawna Grosskopf, Dimitris Margaritis, and William L. Weber

183

6. Productivity Measurement in Sectors with Hard-to-Measure Output Kim Zieschang

211

7. Productivity Measurement in the Public Sector W. Erwin Diewert

241

8. Productivity Measurement and the Environment Finn R. Førsund

287

vi Contents

PA RT I I I M IC ROE C ON OM IC S T U DI E S 9. Productivity and Financial Performance Emili Grifell-Tatjé and C. A. Knox Lovell 10. Business Model Innovation and Replication: Implications for the Measurement of Productivity Roberto Garcia-Castro, Joan Enric Ricart, Marvin B. Lieberman, and Natarajan Balasubramanian 11. The Labor Productivity of Family Firms: A Socioemotional Wealth Perspective Shainaz Firfiray, Martin Larraza-Kintana, and Luis R. Gómez-Mejía

329

359

387

12. Innovation, Management Practices, and Productivity Mary J. Benner

411

13. Internationalization, Innovation, and Productivity Bruno Cassiman and Elena Golovko

437

14. Effect of International Competition on Firm Productivity and Market Power Jan De Loecker and Johannes Van Biesebroeck 15. Efficiency Measures in Regulated Industries: History, Outstanding Challenges, and Emerging Solutions Laurens Cherchye, Bram De Rock, Antonio Estache, and Marijn Verschelde 16. Theory, Techniques, and Applications of Regulatory Benchmarking and Productivity Analysis Per J. Agrell and Peter Bogetoft

463

493

523

PA RT I V M AC ROE C ON OM IC S T U DI E S 17. Productivity and Welfare Performance in the Public Sector Mathieu Lefebvre, Sergio Perelman, and Pierre Pestieau

557

18. Measuring Productivity Dispersion Eric J. Bartelsman and Zoltan Wolf

593

Contents vii

19. Decomposing Value-Added Growth into Explanatory Factors W. Erwin Diewert and Kevin J. Fox 20. The World KLEMS Initiative: Measuring Productivity at the Industry Level Dale W. Jorgenson

625

663

21. Productivity and Substitution Patterns in Global Value Chains Marcel P. Timmer and Xianjia Ye

699

22. The Industry Sources of Productivity Growth and Convergence Robert Inklaar

725

23. Productivity and Economic Development Hak K. Pyo

753

24. The Productivity of Nations 781 Oleg Badunenko, Daniel J. Henderson, and Valentin Zelenyuk Author Index Subject Index

817 837

Contributors

Per J. Agrell, Université Catholique de Louvain Oleg Badunenko, University of Portsmouth Natarajan Balasubramanian, Syracuse University Bert M. Balk, Rotterdam School of Management, Erasmus University Eric J. Bartelsman, Vrije Universiteit Amsterdam Mary J. Benner, University of Minnesota Peter Bogetoft, Copenhagen Business School Bruno Cassiman, IESE Business School, University of Navarra Laurens Cherchye, Katholieke Universiteit Leuven Jan De Loecker, Princeton University Bram De Rock, Université Libre de Bruxelles W. Erwin Diewert, University of British Columbia Lucy P. Eldridge, US Department of Labor Antonio Estache, Université Libre de Bruxelles Rolf Färe, Oregon State University Shainaz Firfiray, University of Warwick Finn R. Førsund, University of Oslo Kevin J. Fox, University of New South Wales Roberto Garcia-Castro, IESE Business School, University of Navarra Elena Golovko, Tilburg University Luis R. Gómez-Mejía, Arizona State University Emili Grifell-Tatjé, Universitat Autònoma de Barcelona Shawna Grosskopf, Oregon State University

x Contributors Daniel J. Henderson, University of Alabama Robert Inklaar, University of Groningen Dale W. Jorgenson, Harvard University Martin Larraza-Kintana, Universidad Pública de Navarra Mathieu Lefebvre, Université de Strasbourg Marvin B. Lieberman, Anderson School of Management, UCLA C. A. Knox Lovell, University of Queensland Dimitris Margaritis, University of Auckland Sergio Perelman, Université de Liège Pierre Pestieau, Université de Liège Hak K. Pyo, Seoul National University Joan Enric Ricart, IESE Business School, University of Navarra R. Robert Russell, University of California, Riverside Robin C. Sickles, Rice University Chris Sparks, US Department of Labor Jay Stewart, US Department of Labor Marcel P. Timmer, University of Groningen Johannes Van Biesebroeck, Katholieke Universiteit Leuven Marijn Verschelde, IÉSEG School of Management William L. Weber, Southeast Missouri State University Zoltan Wolf, CES, US Census Bureau Xianjia Ye, University of Groningen Valentin Zelenyuk, University of Queensland Kim Zieschang, University of Queensland

T h e Ox f o r d H a n d b o o k o f

P RODU C T I V I T Y A NA LYSI S

Pa rt I

E DI TOR S’ I N T RODU C T ION

Chapter 1

Overview of Productivit y A na lysi s History, Issues, and Perspectives Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles

1.1. Introduction Our objective in this chapter is to provide an overview of some important aspects of productivity analysis, many of which are addressed in subsequent chapters in this Handbook. In section 1.2 we stress the economic significance of productivity growth. In subsections 1.2.1 and 1.2.2 we focus on the impact of productivity growth on business financial performance and on the growth of the aggregate economy, the two being linked by the fact that successful businesses grow, and their expansion drives growth in the aggregate economy. At each level, productivity growth has been a historically important driver of performance, although its degree of importance has varied with trends in other potential drivers and with external circumstances. In subsection 1.2.3 we assume that aggregate productivity growth occurs and ask whether this is sufficient for an improvement in economic welfare, a much broader concept than that of economic output such as gross domestic product (GDP). This leads us into the literature directed at the increasingly popular but stubbornly elusive concepts of social progress and inclusive growth. In section 1.3 we explore definition, quantification, and implementation, the procedures through which productivity measures are obtained. In subsection 1.3.1 we define alternative measures of productivity and its rate of change, suggesting some properties that these measures might be asked to satisfy. In subsection 1.3.2 we review two approaches to quantifying productivity change: one, which we call calculation,

4 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles based exclusively on quantity and price data, and the other, which we call estimation, based on quantity and price data augmented by economic theory. In subsection 1.3.3 we consider some implementation issues confronting statistical agencies responsible for constructing and disseminating productivity and related measures of economic activity. In section 1.4 we introduce productivity dispersion among producers. Dispersion matters because aggregate productivity is inversely correlated with the extent of disaggregate dispersion. In subsection 1.4.1 we review the evidence, which shows productivity dispersion to be widespread. In subsection 1.4.2 we introduce productivity dynamics, which considers two possible consequences of productivity dispersion. In one scenario, market forces generate a reallocation of resources away from productivity laggards toward productivity leaders that narrows dispersion and raises aggregate productivity. In the other scenario, barriers to the working of market forces or other factors allow dispersion to persist. In section 1.5 we consider some forces, both internal and external to business, which contribute to productivity and its dispersion. In subsection 1.5.1 we consider technology- based drivers, for which we need information on the underlying production technology. We define productivity change in terms of the technology, and we decompose productivity change into its technology-based drivers, historically the most significant of which has been technical progress, which is inferred from outward shifts in production technology. In subsection 1.5.2 we analyze organizational and institutional drivers, the former being internal to the business and the latter being external to the business. The organizational drivers revolve around management and its practices, including human resource practices, and technology adoption strategies. The institutional drivers include various features of the business’s operating environment that can enhance, as well as impede, productivity. Many of these drivers, such as the regulatory environment, are amenable to public policy intervention. The lengthy list of references that follows is meant to serve as a readers’ guide to the topics we discuss, and to encourage interdisciplinary reading.

1.2. The Significance of Productivity Growth 1.2.1. The Microeconomic Significance of Productivity Growth Productivity growth enhances business financial performance, but its contribution is concealed by conventional financial statements expressed in current prices. This motivated Davis (1955), writing during a period of sharply rising price levels following World War II, to develop a common-price accounting framework, which he called

Overview of Productivity Analysis 5 “productivity accounting.” This framework, in conjunction with the conventional current-price accounting framework, enabled him to isolate the contribution of productivity change from that of price change to business profit change, and it can be extended to alternative indicators of business financial performance. Kendrick and Creamer (1961) and Kendrick (1984, 52) stressed the microeconomic significance of productivity growth, with Kendrick claiming that “. . . over the long run, probably the most important factor influencing profit margins is the relative rate of productivity advance. . . . In the short run, the effects of productivity trends may be obscured.” The short-run phenomena include price movements, as Davis noted, and both Kendrick and Creamer, and Kendrick, used a variant of Davis’s productivity accounting to separate the impacts of productivity advance from those of price changes. Much later, in his survey of the determinants of productivity, Syverson (2011, 327) embellished Kendrick’s claim, referring to what he called a “robust” finding in the literature, namely that “. . . higher productivity producers are more likely to survive than their less efficient industry competitors . . .” and consequently productivity “. . . is quite literally a matter of survival for businesses.” However, as Davis and Kendrick noted, productivity change is not the only driver of change in business financial performance, particularly in the proverbial short run; price change matters as well, as minerals companies around the world have learned, to their joy and despair, since the year 2000. Davis (1955), Eldor and Sudit (1981), and Miller (1984) developed models capable of decomposing change in business financial performance into the separate impacts of quantity changes and price changes, and to quantifiable changes in the operating environment as well. Grifell-Tatjé and Lovell (2015) provide a theory-based approach to identifying the sources of change in business financial performance, an approach they summarize and extend in Chapter 9 of this Handbook. They first decompose profit change into quantity change and price change components. They then decompose the quantity change component into a productivity effect and a quantity margin, or replication, effect, and they decompose the price change component into a price recovery effect and a price margin, or inflation, effect. This identifies four distinct drivers of business financial performance. While the subject of this Handbook is productivity, a rich literature, highlighted perhaps by Winter and Szulanski (2001), stresses the significance of expansion through replication as a business strategy that has become the main driver of the financial performance of businesses such as Walmart, Starbucks, and fast food chains. Garcia-Castro, Ricart, Lieberman, and Balasubramanian illustrate the value of the replication effect for Southwest Airlines in Chapter 10 of this volume, although their replication effect differs analytically from that of Grifell-Tatjé and Lovell. However, if profit is treated as the return to capital and expensed, as in national income accounting and in the investigation of the sources of US productivity growth by Jorgenson and Griliches (1967), and at the individual business level as advocated by Davis (1955), Kendrick and Creamer (1961), and Eldor and Sudit (1981), then the two margin effects disappear, leaving just two drivers of financial performance, productivity change and price recovery change, and replication is not an issue.

6 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles Lawrence, Diewert, and Fox (2006) develop an alternative analytical framework within which to analyze the impact of productivity change on change in business financial performance. They express business profit change in ratio form, and decompose it into the product of three components: productivity change, price recovery change, and change in the size of the business. Their analytical framework distinguishes primary inputs from intermediate inputs, defines business size in terms of its primary inputs, and treats profit as gross operating surplus, the difference between revenue and intermediate input expense. They apply their framework to the financial performance of Telstra, Australia’s largest telecommunications provider, over the 1984–1994 decade, and find productivity growth to have been the sole driver of growth in Telstra’s gross operating surplus, which was depressed by declines in the real prices of its telecommunication services. The fact that the productivity effect and the price effect work in opposite directions is not an uncommon finding, especially in extractive industries during boom and bust cycles; Topp, Soames, Parham, and Bloch (2008) recount the Australian mining experience. From an analytical perspective, and assuming fixed prices, productivity growth leads either to an increase in output (and therefore revenue) per unit of input (and therefore cost), or to a reduction in input use (and therefore cost) per unit of output (and therefore revenue). In either case, productivity growth is lucrative, on several indicators: it increases unit revenue or it reduces unit cost; it increases profit and it increases profitability, the ratio of revenue to cost. Particularly when outputs are not under management control, it is appropriate to explore the impact of productivity change exclusively on cost or unit cost. An example is provided by the financial performance of regulated utilities that face exogenous demand and are allotted by the regulator a share of the profit they generate from cost-reducing productivity improvements. In Chapter 15 of this Handbook, Cherchye, De Rock, Estache, and Verschelde trace the evolution of the use of frontier-based techniques in the analysis, and incentive- based regulation, of the performance of infrastructure industries. They provide a list of seven infrastructure industries that have been subjected to various forms of frontier- based incentive regulation in 21 countries. Of special interest is the attention they pay to the policy arenas within which incentive regulation has developed; the interaction among academics, regulators, operators, and policymakers has a long and continuing history. Against this background they propose a structural approach to performance measurement in regulated industries based on Economic objectives of the participating agents, the structure of production Technology, and Challenges associated with information asymmetry (ETC). No component of ETC is known, of course, and each must be specified by the analyst, presumably constrained by the prerequisites enumerated by Agrell and Bogetoft in Chapter 16 of this volume. Agrell and Bogetoft survey theory and techniques, predominantly frontier-based, used in regulatory benchmarking. They begin with a demanding list of prerequisites for regulatory benchmarking, a list that is relevant not just to regulatory benchmarking but to virtually all empirical economic analysis. They continue with an equally demanding list of elements underpinning the use of frontier techniques in regulatory

Overview of Productivity Analysis 7 benchmarking. In their dynamic yardstick model, the regulator compensates a regulated firm under evaluation with a lump sum payment minus actual cost plus a fraction of the frontier-based estimated cost savings, the latter defined as the difference between a cost norm calculated from a super-efficiency cost-frontier model based on all regulated firms, excluding the regulated firm under evaluation and the regulated firm’s actual cost. The authors illustrate the workings of their models with several samples of European transmission and distribution system operators.1 Productivity growth also increases such commonly used financial performance indicators as return on assets (ROA) and return on equity. This extension is important because these two financial performance indicators feature prominently in the business press reporting of corporate financial performance. They also serve extensively in the academic literature, both as dependent variables in regressions attempting to identify significant drivers of business financial performance, such as ownership and governance structure, and also as independent variables in regressions attempting to quantify the impact of business financial performance on firm growth. Illustrating the latter line of research, Coad (2007) and Bottazzi, Secchi, and Tamagni (2008) provide empirical evidence bearing on the evolutionary principle of “growth of the fitter,” derived from the works of Schumpeter (1942), Alchian (1950), and Nelson and Winter (1982). The principle posits that relatively fit firms, being defined as being either more productive or more profitable than less fit firms, subsequently exhibit faster growth, thereby increasing their market share at the expense of less fit firms. Empirical support for both versions of the hypothesis is weak.

1.2.2. The Macroeconomic Significance of Productivity Growth Microeconomic productivity growth aggregates to macroeconomic productivity growth. Fabricant (1961, xxxv) stressed the macroeconomic importance of productivity growth in stating that “[h]igher productivity is a means to better levels of economic well-being . . .” and that productivity “. . . affects costs, prices, profits, output, employment and investment, and thus plays a part in business fluctuations, in inflation, and in the rise and decline of industries.” Kendrick (1961, 3) concurred, claiming that productivity growth has generated “. . . a large gain in the plane of living . . .” and an increase in “. . . the quality and variety of goods . . . while increasing provision was made for future growth. . . .” Many scholars have studied the contribution of productivity growth to output growth. Schmookler (1952) attributed “about half ” of growth in US output to productivity growth (“increased efficiency of resource use”) over the period 1869–1938.2 Kendrick (1961) attributed “about half ” of growth in US output to productivity growth over the period 1899–1957. Denison (1962, 1974) attributed less than half of growth in US output to productivity growth over the periods 1929–1957 and 1950–1962. Something else has been driving US output growth, namely, input growth. The role of input growth in the studies

8 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles cited in the preceding is apparent; in other studies it has been open to debate. The most- studied example may be the East Asian Miracle, in which rapid output growth in the regional economies since 1960 was, or was not, due primarily to rapid input growth, rather than impressive productivity growth. Hsieh (2002) and Felipe and McCombie (2017) summarize the debate, with the former focusing on the use of quantities or prices, and the latter focusing on the use of values related through an accounting identity, rather than quantities related through technological relationships, and concluding that the debate was “much ado about nothing.” Studies of the contribution of productivity growth to growth in output per person are also numerous. The relevance of output per person is emphasized by Gordon (2016), who interprets output per person as “the most accessible,” though still flawed, definition of the standard of living. We return to the flaws in section 1.4. Gordon examines trends in output per person in the United States through the very long period 1890–2014. For the period 1890–1920, Gordon finds productivity growth to have accounted for approximately one-fourth of growth in output per person. This ratio tripled during the 1920–1970 period that he calls the “Great Leap Forward,” and declined to approximately one-third during the 1970–2014 period, which led him to conclude that some inventions are more important than others. Crafts and O’Rourke (2013) study a number of countries over the very long period 1870–2007 and find two growth spurts, the larger, their 1950–1973 “Golden Age,” roughly coinciding with the second half of Gordon’s “Great Leap Forward,” and the smaller during 1990–2007. Bergeaud, Cette, and Lecat (2016) also find two productivity waves over the very long period 1890–2012 in a sample of 13 advanced economies, the larger generally coinciding with Gordon’s “Great Leap Forward” during 1920–1970 and the smaller occurring after 1995 and primarily in the United States. Both Gordon (2016) and Bergeaud, Cette, and Lecat (2016) emphasize that their respective larger productivity waves occurred long after the innovations that created them, such as the internal combustion engine, the assembly line, and electric light and power. Their emphasis on lags recalls David’s (1990) reaction, from the perspective of an economic historian, to Solow’s (1987) famous “productivity paradox”: avoid the “pitfall of unrealistic impatience,” because it takes time for innovations, general-purpose technologies in particular, to diffuse through economies. Nonetheless, van Ark (2016, 4) reminds us that the productivity paradox is alive and well in the new digital economy, although he heeds the warnings of the economic historians by noting that we have not yet progressed from the installation phase to the deployment phase, “. . . when the new technological paradigm will have been widely diffused and will have become common practice across organizations, enabling its full potential in terms of economic and business growth, productivity, and profitability.” Diffusion has become a recurring theme in much of the productivity literature. In his “biography” of productivity, Hulten (2001) tracks output per person and total factor productivity (TFP) in the United States from 1779 to 1997. He finds a very different historical pattern, with the contribution of productivity growth to growth in output per person having varied widely by decade, with contributions ranging from

Overview of Productivity Analysis 9 6.2% (1949–1959) to 161.7% (1859–1869). Earlier studies include Abramovitz (1956), who attributed 87% of growth in output per person in the United States to productivity growth from the 1869–1878 decade to the 1944–1953 decade, a share exactly the same as the more celebrated share calculated by Solow (1957) for growth in output per unit of labor (rather than per person, and the distinction can be economically significant) over the 1909–1949 period. The other driver of growth in output per person in these studies has been capital deepening, an increase in the capital intensity of production, although its role has been far smaller than that of input growth in studies of output growth. One way to increase the role of input growth is to expand the list of inputs, perhaps by decomposing existing inputs. This is one approach followed in the new growth literature initiated by Romer (1986), Lucas (1988), and Mankiw, Romer, and Weil (1992), to name three of the more influential contributions. Romer emphasizes knowledge gained from an endogenous research technology, and the externalities associated with such knowl edge. Lucas emphasizes human capital accumulated through education and learning by doing, and the externalities it creates. Mankiw, Romer, and Weil retain the elements of Solow’s neoclassical growth model, and also emphasize human capital accumulation and the externalities it generates. Somewhat later, Hall and Jones (1999) augment physical and human capital with a host of endogenous institutions and government policies. Easterly and Levine (2001) stress the primacy of productivity growth over factor accumulation, and also emphasize the role of national policies and truly exogenous factors such as economic geography as potential growth determinants. Each of these factors, from human capital to national policies, continues to influence empirical studies of economic growth, and each appears in various chapters in the Handbook. The cost-reducing impact of resource-saving improvements in technology mentioned in subsection 1.2.1 also aggregates to the macroeconomic level, where it has been used to analyze competitiveness among nations. The Organisation for Economic Co-operation and Development (OECD)3 views a country’s unit labor cost as “. . . a broad measure of (international) price competitiveness.” Unit labor cost, the ratio of labor compensation to output, can be converted to the ratio of the wage rate to labor productivity. This conversion demonstrates that a country’s competitiveness is dampened by increases in labor’s wage (which may also be influenced by exchange rate movements) and enhanced by increases in labor productivity. Even within the European Union, with no exchange rate effect to complicate matters, competitiveness varies. German unit labor cost remained unchanged between 2001 and 2014 because increases in labor productivity compensated for increases in labor compensation per hour worked. Italian unit labor cost, on the other hand, increased by 2.3% per annum because labor productivity remained unchanged.4 This labor-oriented concept of competitiveness extends easily to unit cost and multifactor productivity, although difficulties in measuring capital hamper its widespread application. There is an alternative approach to evaluating the contribution of productivity change to change in a country’s real income, with an analytical framework that inspired the framework developed by Lawrence, Diewert, and Fox (2006) for evaluating

10 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles the contribution of productivity change to change in business financial performance that we discuss in subection 1.2.1. Based on the principle that an increase in a country’s terms of trade (the ratio of an index of its export prices to an index of its import prices) should have a qualitatively similar impact on its real income as an increase in its productivity, Diewert and Morrison (1986) developed an analytical framework in which change in a country’s real income is decomposed into the product of the value of productivity change, the value of terms of trade change, which in principle becomes more important as countries open up to international trade, and the value of primary input change. Diewert (2014) applies this model to aggregate US data over 1987–2011, and finds productivity growth and increases in primary inputs to have been the main drivers of growth in real income, with changes in the terms of trade contributing virtually nothing.

1.2.3. The Significance of Productivity Growth for Social Economic Progress Baumol, Blackman, and Wolff (1989, 9–10) claim that “[i]n terms of human welfare, there is nothing that matters as much in the long run” as productivity (emphasis in the original). Prefacing their brief survey of the long run, they stress that “. . . the magnitude of the changes [in living standards] is so great that they resist comprehension.” Notice that they write broadly of human welfare and living standards, not narrowly of some measure of national income, or even of national income per capita. Productivity growth increases aggregate output, but what is its impact on broader measures of national well-being that might incorporate leisure, environmental quality, or the distribution of the national income as well as its magnitude? This question was the subject of a lively debate that occurred subsequent to the Great Depression concerning what constituted “economic progress.” One group, led by Ayres (1944), defined economic progress narrowly as productivity-driven output growth. The other group, including Clark (1940) and Davis (1947), defined economic progress more broadly to augment productivity-driven output growth with rapid re-employment of resources (primarily labor) displaced by resource-saving productivity improvements, a minimum inequality in the distribution of the income created by productivity growth, the imposition of minimal social costs, and various other criteria. The Economist (2016b) provides a good historical introduction to “the machinery question,” which refers to productivity growth that displaces labor by machinery, and which it traces back to the nineteenth- century writings of David Ricardo. The subject of economic progress resurfaced in the 1970s with the Nordhaus and Tobin (1972) “measure of economic welfare,” which adjusts aggregate output for environmental and other impacts, and has resurfaced again recently, most notably in the writings of Stiglitz, Sen, and Fitoussi (2009) on social progress, the OECD (2014, 2016a) on inclusive growth, and Gordon (2016, Chapters 1 and 18) on the conceptual gap between conventional measures of aggregate output and the standard of living, and the

Overview of Productivity Analysis 11 “headwinds” threatening to expand the gap in the near future. Both concepts include the impacts of productivity growth on leisure, the environment, and income inequality, as well as on output. The basic idea is unchanged: the contribution of productivity growth to output, which is measurable, can exceed or fall short of its contribution to some measure of national well-being, which is far more difficult to measure. We call change in this measure of national well-being social economic progress, combining the early concept of economic progress with the recent concept of social progress.5 In Chapter 23 of this volume, Pyo explores the complex relationships linking productivity growth and economic development, which he distinguishes from both economic growth and inclusive growth, although it has much in common with the latter concept. The distinction turns largely on the impact of productivity growth on inequality in the distributions of income and wealth, and on the extent of poverty. Although productivity, distribution, and poverty are empirically correlated, the strength of the correlation is contextual, and causality is difficult to establish. In addition, inequality and poverty are also influenced by geography, demography, and the quality of institutional arrangements such as the security of property rights, barriers to investment in physical and human capital, access to credit, and the structure (and efficacy) of the tax system. Pyo provides a detailed survey of theoretical developments and empirical evidence on the productivity–development connection. Whatever the name, the challenge is to develop business and public policies that counter Gordon’s headwinds. Banks (2015), Hsieh (2015), and the OECD (2015a, 2015b) propose a suite of public policies to promote productivity. The list is long and unsurprising, including labor and product market reforms that would lower barriers to entry and exit; promoting business investment in research and development (R&D), information and communications technology (ICT), and other forms of knowledge-based capital; promoting reforms to the financial sector that would increase access to, and reduce the cost of, capital; promoting investment in public infrastructure; lowering tariff and nontariff barriers to promote cross-border trade and investment; promoting the transfer of resources from public to private ownership; and adopting policies that would encourage the transfer of resources from the informal sector to the formal sector, particularly in poor countries where the informal sector is large. The OECD (2015a, 2015b) also recognizes that some pro-growth policies have unintended adverse consequences for the environment and economic equality, to which we now turn.

1.2.3.1. The Environment When the environment is involved, two issues arise in relation to productivity growth. Both issues involve the question: Productivity growth increases national income, but if it has adverse environmental impacts, what is its contribution to social economic progress? First, suppose production activity has adverse environmental impacts, such as air and water pollution, and the natural environment serves as a receptacle for the disposal of pollutants, perhaps but not necessarily because disposal is free, or priced beneath marginal abatement cost. Productivity measures that exclude these impacts generally differ

12 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles from measures that incorporate them. On what grounds do we prefer one to the other? And how might environmental impacts be incorporated into a holistic model of productivity change? Several writers, from Førsund (2009) and Lauwers (2009) through Dakpo, Jeanneaux, and Latruffe (2016) and the studies contained in Kumbhakar and Malikov (2018), critically survey existing models and find them lacking for their neglect of the materials balance condition, which states, in Førsund’s words in Chapter 8 of this volume, that “[i]f all the material inputs into an activity are not embedded in the products the activity is set up to deliver, then the difference must be contained in residuals discharged to the environment.” Dakpo, Jeanneaux, and Latruffe develop a model that incorporates the generation of residuals in a way that respects the materials balance condition, and facilitates productivity measurement incorporating residuals. In Chapter 8, Førsund surveys the literature devoted to incorporating environmental impacts in productivity modeling, critically evaluating the standard approach based on the concept of weak disposability of residuals and sketching a new approach that dispenses with the weak disposability assumption. The preferred new approach is influenced by the work of Ragnar Frisch (1965) and is still being developed. Its four key features are the following: (i) a decomposition of the output vector into intended outputs and residuals that may pollute the environment; (ii) a decomposition of the input vector into non-materials inputs used to produce the intended outputs and materials inputs that contribute to the production of the intended outputs and also are responsible for generation of the residuals; (iii) satisfaction of the materials balance condition defined earlier; and (iv) a multi-equation system derived from a pair of linked production technologies, one for the intended outputs and the other for the residuals. After developing the model, Førsund surveys alternative regulatory approaches to environmental impacts. In this context he dismisses as “Panglossian” the Porter (1991) hypothesis, which states, loosely, that sufficiently strict environmental regulation may induce polluting firms to innovate to such an extent that pollution diminishes and, simultaneously, profit increases, making environmental regulation a win-win strategy. Albrizio, Kozluk, and Zipperer (2017) provide empirical evidence bearing on the impacts of environmental regulation on productivity. They use data from 19 OECD countries during 1990–2010 to develop an index of environmental policy stringency, which they relate to multifactor productivity at three levels of aggregation: economy, industry, and firm. At the economy level they find a negative announcement effect that is offset within three years of the imposition of more stringent environmental regulations. At the industry level they find a temporary productivity boost from an increase in regulatory stringency for technologically advanced country-industry pairs, with the effect diminishing with declining levels of advancement. At the firm level, the focus of the Porter hypothesis, they find that only the most technologically advanced firms enjoy a productivity boost from more stringent environmental regulations; the less productive third of all firms experience productivity declines. Second, suppose natural resources such as coal, petroleum, or natural gas are used as inputs in the production process. A number of issues have been raised concerning natural resource depletion, including (i) developing environmental accounts that

Overview of Productivity Analysis 13 would incorporate natural resource stocks, including land and sub-soil assets, that would support environmental productivity accounting; (ii) modeling and measuring the impacts of changes in natural resource stocks on estimates of productivity change; and (iii) addressing the intergenerational issue of whether these stocks are being depleted in some optimal sense. The first issue is addressed by Førsund in Chapter 8; by Nordhaus and Kokkelenberg (1999), who trace the history to date of national income and product accounts, augmented national accounts, and integrated accounts in the United States; and by Bartelmus (2014), who refers to conventional measures of aggregate output as “environmentally and socially blind,” and who relates progress, or lack thereof, in the development of an integrated System for Environmental and Economic Accounting (SEEA) at the United Nations. Bartelmus (2015) continues by noting that SEEA incorporates the interaction between the economy and natural resource depletion, and bemoans the current lack of a system of accounts that incorporates the interaction between the economy and both natural resource depletion and environmental degradation. The second issue is addressed by Topp and Kulys (2014), who show that declines in quality-adjusted stocks of (renewable and nonrenewable) natural resources reduce estimated rates of productivity growth when the resource is both an input to and the output of a production activity. Examples include fishing and mining, in which, as resource depletion makes constant quality stocks more difficult to access, additional amounts of other inputs are required to produce a given amount of fish or mineral output, and measured productivity declines. Brandt, Schreyer, and Zipperer (2017) use a standard growth accounting framework to show that, when natural capital is not included among the inputs, input growth is under-(over-)estimated by the traditional measure when the natural input grows faster (slower) than the included inputs, and consequently productivity growth is biased upward (downward). They illustrate their analysis with data from the OECD Productivity Database, augmented with natural capital data sourced from the World Bank. The third issue is ongoing, and recalls the famous 1980 wager, made in the wake of the “Limits to Growth” movement, between biologist Paul Ehrlich and economist Julian Simon on whether the real price of a bundle of resources would rise or fall between 1980 and 1990. The real price of the bundle of resources fell, and Simon won the bet.

1.2.3.2. Inequality The second prominent strand in the economic progress and inclusive growth literature concerns the distribution of the fruits of productivity growth. The general question remains: Productivity growth increases national income, but if it increases inequality in the distribution of income, what is its impact on social economic progress? Gini coefficients have increased around the world, leading to the hypothesis that productivity growth exacerbates inequality in the distribution of income. Credible evidence must distinguish between correlation and causation, and OECD (2014, 2015b, 2016a) appears to assert causation, identifying several drivers of rising income inequality, the most important being (i) skill-biased technical progress that has caused

14 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles increasing dispersion in wages and salaries and displacement of less-skilled labor; (ii) regulatory reforms that have increased product and labor market flexibility also have contributed to increased wage inequality; and (iii) rising shares of nonwage income from capital have increased income inequality. Before exploring these three drivers, however, it is worth noting that the OECD6 also believes that productivity growth has slowed since the global financial crisis (GFC), while inequality of income and opportunity has been growing. This admittedly short-term trend suggests a negative correlation between the two, which in turn offers cause for optimism if productivity growth reverts to trend. Nonetheless, the McKinsey Global Institute (2016), in a study of 25 advanced economies, find that two-thirds of households were in segments of the income distribution whose real incomes were flat or declined in the decade to 2014, and conclude that today’s younger generation is at risk of ending up poorer than their parents. They suggest enactment of government tax and regulatory and welfare policies, but they also recommend business policies aimed at increasing productivity and growth, and reducing inequality. Recommended business policies include adoption of existing best practices and pursuit of technological and operational innovations that expand best practices.7 Rising Gini coefficients are a cause for concern, but they need to be interpreted carefully. We have noted that the OECD has documented rising Gini coefficients within OECD economies. Lakner and Milanovic (2016) also document rising Gini coefficients within each of five world regions, with three more regions having incomplete data. However, they also find a declining global Gini coefficient, the ostensible conflict being attributable to the rapid growth of per capita incomes in still-low-income China and India. Capital-skill complementarity exists, as Griliches (1969) pointed out long ago, and it is by no means a recent phenomenon, as Goldin and Katz (1998) have demonstrated. In addition, improvements in technology tend to be biased, using capital and skilled labor and saving less-skilled labor, as demonstrated by Krussell, Ohanian, Ríos-Rull, and Violante (2000). This combination of bias and complementarity motivates the first driver of rising income inequality mentioned in the previous paragraph. Brynjolfsson and McAfee (2014) express optimism for skilled labor but pessimism for less-skilled labor. Frey and Osborne (2013) paint a gloomy picture, predicting that computerization will put at risk of displacement nearly half of US employment, the half performing manual and/or routine tasks, a prediction that has earned them the moniker “techno- pessimists.” Autor (2015) and Mokyr, Vickers, and Ziebarth (2015) review the history of automation, unemployment, and re-employment, and take a different view. While automation, through computerization, robotics, and artificial intelligence, does displace labor, it also complements labor, eventually raising output, the demand for (different types of) labor, and employment. Autor usefully distinguishes jobs from tasks, noting that while manual and routine tasks disappear, many jobs that require multiple tasks that cannot easily be unbundled survive. The OECD (2016a) traces the second driver to the fact that labor market liberalization creates nonstandard forms of work having relatively low wages and benefits, thereby putting upward pressure on inequality. The Economist

Overview of Productivity Analysis 15 (2016a) emphasizes the third driver, arguing that the productivity dividend is being hoarded, ending up as business retained earnings rather than being invested in growth- and employment-boosting activities, an observation that is repeated frequently. In a fascinating juxtaposition of rock & roll and economics, Krueger (2013)8 provides a wealth of secondary information on income distribution in the music industry and in the US economy. He views the music industry as a microcosm of the economy, buffeted by technological changes, scale, luck, and an erosion of social norms, which compress prices and incomes, all of which have contributed to an increasingly skewed artist income distribution. As for the economy, he finds (i) all family income quintiles experienced over 2% annual growth during 1920–1979, but only the top quintile experienced over 1% annual growth during 1979–2011, and the bottom quintile experienced a decline; (ii) the share of income earned by the top 1% has roughly tripled, to 18%, since 1970; (iii) the ratio of CEO to average worker compensation has increased tenfold, to over 200-to-1, since 1970; and (iv) profitable companies pay all employees well, janitors as well as managers, suggesting that growing inequality originates between companies rather than within them. He concludes by expressing concern that rising inequality may have adverse consequences for future economic growth. Song, Price, Guvenen, Bloom, and Wachter (2015) develop an analytical framework that enables them to decompose wage dispersion into between- firm and within-firm wage components. Their empirical findings, obtained from a matched employer-employee data set covering all US firms during 1978–2012, reinforce Krueger’s final point; rising aggregate wage inequality is due exclusively to increasing inter-firm wage dispersion, since within-firm wage dispersion has been stable. The authors do not take the obvious next step of inquiring whether firms paying higher wages are also more productive. Doing so might shed light on the productivity dispersion literature we discuss in section 1.4. In Chapter 9 of this volume, Grifell-Tatjé and Lovell adopt a standard growth accounting framework for measuring productivity change, the mirror image of which is a framework for analyzing the distribution of the value created by productivity change. Productivity growth creates value, which is distributed to consumers (in the form of product price reductions), to labor and other input suppliers (in the form of increased remuneration), and to the business itself (in the form of increased profit, which management allocates to interest, taxes, depreciation, amortization (ITDA) and retained earnings). Product price reductions and input price increase spur growth, as does business investment arising from profit increases. Distribution of the fruits of productivity growth was the principal concern of Vincent (1968), the French public institution CERC (Centre d’Études des Revenus et des Coûts) (1969), and other prominent French scholars who studied distribution at French public firms such as Electricité de France. Grifell-Tatjé and Lovell (2015) survey the extensive French literature. This growth accounting framework supports an analysis of the impacts of price changes on the functional distribution of income, which in turn forms the basis of an analysis of trends in the inequality of the functional distribution of income that can address the drivers identified by OECD and the hoarding issue raised by The Economist.

16 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles We conclude the discussion of income distribution by acknowledging that the income distribution dual to the productivity model characterizes the functional distribution of income among groups who perform a productive service that contributes to value creation, whereas much of the income inequality evidence pertains to the size distribution of income among groups of individuals, income deciles for example, regardless of what, if any, function they perform. The US Bureau of Labor Statistics (BLS)9 reports that labor’s share of value added in the US private business sector has declined from 0.67 to 0.62 from 1987 to 2014, a decline of nearly 7%, which documents a declining labor share in the functional distribution of income, but says very little about rising inequality in the size distribution of income in the United States. At the other extreme, the OECD (2014, 2016a, 2016b) provides extensive documentation of rising inequality in the size distribution of income in most OECD countries using a Gini coefficient based on “equivalised household disposable income,” which, conversely, says very little about trends in the functional distribution of income within countries. Garvy (1954) attempted to reconcile the differences between the two concepts of income distribution, and Fixler and Johnson (2014) search for evidence on the size distribution of income in the US national accounts. Additional research that would reduce, or at least clarify, the gap between the two measures of income distribution would enlighten the inequality debate.

1.3. Productivity Measurement Evaluating the significance of productivity growth, as in section 1.2, requires defining productivity and then quantifying it, which in most instances can be achieved in either of two quite different ways. We consider how to define productivity in subsection 1.3.1, and we consider two approaches to quantification of productivity change in subsection 1.3.2. In subection 1.3.3 we consider how productivity measurement is implemented by statistical agencies, and we discuss two important measurement problems that must be addressed.

1.3.1. Definitions In general, productivity is defined as a ratio of output to input Y/X, with Y an output aggregator and X an input aggregator. An index of productivity change can be expressed as the ratio of an output quantity index Yt+1/Yt to an input quantity index Xt+1/Xt, with t a time indicator, or as a rate of growth GY/X = GY – GX, with G a growth rate. Setting X = X(L) generates a popular partial productivity measure Y/X(L), with X(L) either a scalar (e.g., hours worked) or a scalar-valued aggregate of various characteristics of labor. Setting X = X(K, L) in a value added context or X = X(K, L, E, M, S) in a gross output context generates a pair of multifactor productivity (MFP) measures Y/X(K, L)

Overview of Productivity Analysis 17 and Y/X(K, L, E, M, S). In each case, an index of productivity change is defined as in the preceding. Balk (2009) and OECD (2001a) explore the relative merits of the value added and gross output MFP indices. As Eldridge, Sparks, and Stewart note in Chapter 3 of this Handbook, the BLS publishes KLEMS-based MFP tables for US manufacturing and non-manufacturing sectors and industries for the period 1987–2014.10 In Chapter 20 of this volume, Jorgenson provides an introduction to the World KLEMS Initiative and its regional components for making international productivity comparisons at the industry level. Jorgenson relates the history of the Initiative, and provides details on the KLEMS framework for productivity measurement, including procedures for developing constant quality indices of the primary labor and capital inputs. A critical component of the Initiative, not relevant to intra-country data construction exercises, is the use of purchasing power parities, as distinct from market exchange rates, to link international currencies. He provides an empirical application of the Initiative with an industry-level productivity comparison of Japan and the United States. Inklaar uses the EU KLEMS database in his study of productivity and convergence in Chapter 22 of this volume. MFP measures are generally preferred to partial productivity measures, but there are exceptions to the general rule. The most prominent exception occurs when labor is the only measurable input comparable across producers, as is frequently the case with international comparisons. A second exception occurs when the focus is on inequality in the size distribution of income, in which case labor productivity may be the relevant productivity indicator. These two exceptions intersect in OECD (2016b), in which declining labor productivity growth rates are contrasted with rising labor income inequality in both OECD and non-OECD countries. Firfiray, Larraza-Kintana, and Gómez-Mejía offer a third exception in Chapter 11. They study family-controlled firms, in which non-economic objectives are important. These objectives include an emotional attachment to the firm, a desire to maintain control of the firm, and a desire to hand the firm down to future generations, a practice that Caselli and Genaioli (2013) call “dynastic management”; these objectives are collectively referred to as the protection of socioemotional wealth (SEW). The authors develop a framework in which SEW protection may be a significant factor explaining differences in labor productivity between family and nonfamily firms, and among family firms of varying sizes. They hypothesize that varying SEW priorities lead to variations in leadership styles, capital investment decisions, the role of nonfamily managers, and human resource management (HRM) practices, which combine to generate labor productivity dispersion; on productivity dispersion, see also subection 1.4.1, and on HRM practices see also subsection 1.5.2. Grifell-Tatjé and Lovell (2004) offer another motivation for using labor productivity. Following the cooperative literature inspired by the Illyrian firm of Ward (1958) and the Soviet collective farm of Domar (1966), they analyze the dividend-maximizing behav ior of Spanish cooperative financial institutions. The dividend each employee receives consists of a wage plus the share of each employee in profit after taxes and interest. Change in the dividend decomposes into the product of labor productivity change, input deepening change, and price changes, making labor productivity change a driver of change in cooperative financial performance.

18 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles Bryan (2007, 1) proposes an intriguing indirect motivation for the use of labor productivity. He notes that the most valuable assets a firm has are not tangible physical assets but intangible assets such as “. . . the knowledge, relationships, reputations and other intangibles created by talented people and represented by investments in such activities as R&D, marketing and training.” Unlike most writers on intangible capital, Bryan treats this asset as a component of the labor input, rather than of the capital input. This motivates him to propose replacing the popular return on (tangible physical) assets financial indicator with return on employees, because talent embedded in a company’s employees is the ultimate profit driver. It is easy to show, using a modified duPont triangle approach, that profit per employee can be expressed as the product of the conventional profit margin on revenue and the revenue productivity of labor. For more on revenue productivity, see section 1.4. The standard MFP concept can be extended to the concept of dynamic productivity. The essential difference is that in a standard productivity model, resources available in a period are used exclusively to produce final outputs in that period, whether or not the productive capacity of those resources is fully utilized, whereas a dynamic productivity model allows a reallocation of resources through time, with some resources available in a period being withheld from current production and made available for use in producing final outputs in a subsequent period. Another feature of a dynamic productivity model is its incorporation of intermediate inputs; produced outputs in a period can be consumed as final output, as in a standard productivity model, or saved and used as intermediate inputs in a subsequent period. A third feature of a dynamic productivity model is its property of time substitution, which allows firms to choose when to begin and cease production, and how intensely to produce; for example, technical progress encourages firms to delay production. The definition of productivity change remains unchanged as GY/X = GY – GX, although the contents of Y and X are modified to incorporate the elements of dynamic productivity. In Chapter 5 of this volume, Färe, Grosskopf, Margaritis, and Weber build on previous work of Shephard and Färe (1980), and Färe and Grosskopf (1996), to develop a dynamic production technology and to define standard and dynamic performance measures that allow for reallocation of resources through time. They derive a dynamic productivity index, decompose it into measures of dynamic efficiency change and dynamic technical change, and relate dynamic efficiency change to standard efficiency change. They provide an empirical application to 33 OECD countries over the period 1990–2011. A comparison of dynamic and standard productivity indices shows slightly faster dynamic productivity growth, due to faster dynamic technical progress. The authors refer to related research, to which additional empirical work would add value.

1.3.2. Quantification Once productivity is defined, it must be quantified, which requires specification of functional forms in two quite different contexts. In one context, quantities and prices are

Overview of Productivity Analysis 19 observed, and we require a functional form that combines them. In the other context, economic theory constrains the behavior of quantities, or of quantities and prices, and we require a functional form that also incorporates the constraints imposed by theory.

1.3.2.1. Calculation One approach to quantification is through calculation, which involves the use of market prices, or proxies for them, to weight individual output and input quantity changes in the calculation of aggregate output and input quantity indices. This procedure generates price-based empirical productivity indices (the most popular being the asymmetric Laspeyres and Paasche indices, and the symmetric Edgeworth-Marshall, Fisher, and Törnqvist indices). Mills (1932), Fabricant (1940), Kendrick (1961), and many others have calculated productivity change using quantities and prices to construct empirical indices of output, input, and productivity. Mills used Fisher indices, while Fabricant and Kendrick used Edgeworth-Marshall indices. In Chapter 2 of this Handbook, Balk explores empirical quantity and price indices, so named because they can be calculated directly from observable quantities and prices. He specifies a set of desirable properties that empirical quantity and price indices should satisfy, including non-negativity, monotonicity, homogeneity, and units-invariance. He uses quantity and price indices to decompose change in profitability, the ratio of revenue to cost, into the product of an MFP index Y/X and a price recovery index P/W, and he relates this decomposition of profitability change to growth accounting techniques. He discusses the relationship between gross output and value added productivity indices, he relates partial productivity indices to MFP indices, and he discusses aggregation of productivity indices over producers. Empirical index numbers are expressed in ratio form, and are employed throughout the Handbook. However, Bennet (1920) demonstrated that it is also possible to express quantity change, price change, and productivity change in difference form as quantity, price, and productivity indicators. Bennet’s indicators can be expressed as the difference analogue to Fisher’s indices. In Chapter 2, Balk specifies a set of desirable properties that empirical quantity and price indicators should satisfy, analogous to those that empirical quantity and price indices should satisfy; Balk (2008) provides details. He then uses quantity and price indicators to decompose change in profit, the difference between revenue and cost, into the sum of a productivity indicator and a price recovery indicator. In Chapter 9 of this volume, Grifell-Tatjé and Lovell relate productivity indicators, expressed in difference form as value changes, to productivity indices, expressed in ratio form as pure numbers. This relationship has the virtue of translating an index of productivity change to its contribution to a firm’s bottom line, as Davis (1955) first showed.

1.3.2.2. Estimation An alternative way of quantifying productivity growth is through estimation of some underlying technological relationship involving quantity and/or price data, which generates estimates of theoretical productivity indices and indicators. Estimation relies heavily on developments in economic theory that suggest particular functions to be

20 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles estimated, and their properties to be imposed or tested, typically monotonicity, curvature, and homogeneity. Estimation can be based on either econometric techniques popular in economics or mathematical programming techniques popular in management science, and both techniques are utilized in the Handbook. Thus scholars have estimated functional forms for production functions and dual-value functions such as cost, revenue, and profit functions to obtain estimates of theoretical indices of output, input, and productivity. Each of these functions can be extended to frontiers that bound, rather than intersect, the data, thereby providing an additional potential source of productivity change or variation, namely change or variation in the efficiency with which any assumed economic objective is pursued. The econometric approach to frontiers, known as stochastic frontier analysis (SFA), was pioneered by Aigner, Lovell, and Schmidt (1977) and Meeusen and van den Broeck (1977), and the mathematical programming approach, known as data envelopment analysis (DEA), was pioneered by Charnes, Cooper, and Rhodes (1978). Sickles and Zelenyuk (forthcoming) develop a frontier framework for productivity measurement and estimation using both SFA and DEA methodologies and their extensions, and for both primal and dual approaches. In Chapter 4 of this volume, Russell explores theoretical productivity indices, so named because they are defined on well-behaved but unobserved production technology and must therefore be estimated. He discusses the properties that a well-behaved technology satisfies, defines Shephard’s (1953, 1970) distance functions on the technology, and enumerates properties satisfied by distance functions, including monotonicity and homogeneity. He uses distance functions to define the technical efficiency of production, and to define technical efficiency change and technical change, the product of which can be interpreted as a theoretical index of productivity change. He defines and decomposes two different productivity indices named after the Swedish statistician Sten Malmquist. He also defines and decomposes a dual, cost-based, productivity index, which has the virtue of allowing change in allocative, as well as technical, inefficiency to be a driver of productivity change. Productivity growth, whether calculated or estimated, accounts for a variable share of economic growth. Debates among researchers on the primary sources of economic growth and development have often been centered on two basic explanations rooted in the decomposition of economic growth sources: factor-accumulation and productivity- growth components. Kim and Lau (1994), Young (1992, 1995) and Krugman (1994), among others, pointed out that rapid economic growth in the emerging areas of the world such as East Asia was largely explained by the mobilization of resources. Alternative explanations to the neoclassical growth model explain economic growth not only in terms of intensive and extensive utilization of input factors, but also due to factors that impact the degree to which countries can appropriate the productivity potential of world technical innovations. Factors such as governmental industrial policies, trade liberalization policies, and political, religious, and cultural institutions are often viewed as central to the ability of countries to catch up with a shifting world production possibilities frontier.

Overview of Productivity Analysis 21 Exogenous productivity growth was the prevailing modeling assumption until the endogenous growth model put forth by Romer (1986) took hold in the late 1980s. The sources of the endogenous growth, often expressed in a reduced form equation that shifts the production function over time, were typically spillovers of one sort or another. For example, if R is such a variable or set of variables, then the production function can be written as Y = A(R)f(K, L, R). The various possible sources of the spillover differentiate much of the endogenous growth literature, at least at the macroeconomic level. For example, Arrow (1962) emphasized learning by doing. For Romer (1986) the endogeneity came from the stock of research and development. For Lucas (1988) it was the stock of human capital. A major source of post–World War II economic growth has been innovation in the form of technological change. There is, however, another interpretation for the reduced-form endogenous technology term in the modern productivity model, specifically the presence of inefficiency. Suppose one defines the endogenous factor in productivity growth as simply a country’s or firm’s differential ability to loosen the constraints on the utilization of the existing world technology. With this interpretation of endogenous productivity effects, Sickles and Cigerli (2009) show that TFP growth is determined by the efficiency with which the existing technology (inclusive of innovations) is utilized. Production spillovers have important implications for economic growth and for its management. If any type of investment whose gains are not internalized by private agents impacts long-run growth, then there is no unique long-run growth path and thus no so-called golden rule. From a public policy perspective, spillovers provide a clear role for government intervention. Government intervention may take many forms if investment is too low from society’s perspective. Investment tax credits or R&D grants are two traditional forms of government intervention. However, government intervention may also take the form of relaxing constraints on businesses via deregulatory reforms, reduced red tape, private-sector market reforms, or any other aspect of the institutional and political mechanism established in a country and its markets that increase A(R) in the production function. The latter set of external effects can be summed up as governmental actions that reduce constraints, or efficiency-enhancing investments. If one examines the new growth paradigm more closely, it must be recognized that it is indistinguishable empirically from the stochastic frontier model wherein A is an efficiency term. A substantial engine of economic growth has been efficiency change. As pointed out by Abramovitz (1986), Dowrick and Nguyen (1989), and Nelson and Wright (1992), among many others, the major sources of country growth differentials in the developed countries after World War II can be explained by the neoclassical growth model amended to include such endogenous factors as knowledge spillovers, technological diffusion, and convergence to a best-practice production process (Smolny 2000). The “new growth theory” implicitly recognizes the role of efficiency in production. One set of papers that provides an explicit efficiency interpretation of this growth process is Hultberg, Nadiri, Sickles, and Hultberg (1999), Ahn, Good, and Sickles (2000), and Hultberg, Nadiri, and Sickles (2004), who introduce inefficiency into the growth

22 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles process. Of course the standard neoclassical model without explicit treatment of efficiency has been used by many authors in examining growth and convergence. Endogenous growth also has been addressed using formal spatial econometric specifications based on both average production/cost models as well as frontier production/cost models. Models that extend the multiplicative spillover effects by expanding A(R) by framing production in a spatial autoregressive setting in order to address network effects or trade flows among countries have been formulated by Ertur and Koch (2007) and Behrens, Ertur, and Koch (2012). More general stochastic frontier treatments that do not force efficiency on the productive units, whether they are countries, states, or firms, have been introduced by Druska and Horrace (2004) in the cross-sectional setting, and for the panel model in a series of papers by Glass, Kenjegalieva, and Paez-Farrell (2013), Glass, Kenjegalieva, and Sickles (2016a, 2016b), and Han, Ryu, and Sickles (2016).

1.3.3. Implementation Statistical agencies around the world implement the measurement of productivity and related economic variables. They have their choice of calculation or estimation approaches, and they make budget-constrained choices concerning what approach(es) to use, what variables to include, and what sectors of the economy to cover. In Chapter 3 of the Handbook, Eldridge, Sparks, and Stewart discuss how labor productivity and MFP indices are constructed at the BLS.11 The BLS also publishes related data on labor compensation, unit labor costs, and labor’s share, and is engaged in a number of projects designed to expand its range of data products, as outlined in Chapter 3. The OECD (2001a) covers similar ground in much greater detail, and also provides links to a number of national statistical agencies that provide similar services.12 Long ago, Denison (1962, 1974) pointed to a factor that confronts implementation of productivity measurement, namely variation in the quality of inputs. He decomposed growth of the labor input into several sources, including hours worked, the age-sex composition, and educational attainment, and he decomposed growth in the capital input into inventories, structures and dwellings. These adjustments previewed those currently employed at the BLS, which decomposes growth in the labor input into hours and a composition effect (accounting for age, education, and gender), and decomposes growth in the capital input into productive capital stock and a composition effect (accounting for the contributions of information processing equipment, R&D, all other intellectual property products and all other capital services). Over the 1987–2015 period in the US private business and nonfarm business sectors, the two composition effects account for about 40% of total input growth.13 Eldridge, Sparks, and Stewart show the contributions of the two composition effects to labor productivity growth in the private business sector over the 1987–2014 period and several subperiods. Eldridge, Sparks, and Stewart explain the two quality-adjustment procedures employed by the BLS, and the

Overview of Productivity Analysis 23 OECD (2001a) devotes two chapters to the adjustment of labor and capital inputs for variation in various measures of quality. Bushnell and Wolfram (2009) provide an empirical example based on the performance of the operators of five US power plants, demonstrating the importance of variation in the quality of a single key employee, the plant operator, for power plant performance. The plant operator is responsible for the monitoring and control of the combustion process, an integral element of the conversion of potential energy in fuel into electrical energy that also includes processing and monitoring of emissions and other waste products. Using hourly data on fuel burned and power output for individual plant operators, they find a statistically significant positive “operator effect” on fuel efficiency, the ratio of electricity output to fuel input, and they calculate that if all operators at these five plants improved to best practice, fuel cost savings of $3.5 million per year could be achieved. This operator effect is analogous to the “good captain hypothesis” explored by Alvarez and Schmidt (2006) and Wolff, Squires, and Guillotreau (2013) for captains of Spanish and French fishing vessels. This hypothesis asserts that differences in catches among vessels are due to differences in the skill of skippers, rather than to luck and other factors such as weather. Although the analytical framework varies across the three studies, in each study variation in labor quality is not accounted for prior to the empirical exercise. Rather, since quality is difficult or impossible to measure, it is inferred from empirical findings. Power-plant operator and fishing-vessel skipper performance provide ex post measures of variation in the quality of a crucial input. The quality issue is equally relevant on the output side, as Fabricant (1940, 1961) emphasized. The BLS accounts for quality change in outputs in a number of ways, as described in Chapter 3.14 The importance of accurate output measurement was highlighted by the Boskin Commission Report (Boskin et al. 1996), which argued that the US rate of inflation had been overestimated, and consequently the rate of real output growth had been underestimated, by 0.6% per annum prior to 1996 due to a failure to incorporate new outputs and improvements in the quality of existing outputs in a timely fashion. The qualitative impact this has on productivity measurement is clear, and to imagine the quantitative impact, try compounding 0.6% per annum over a generation. Output measurement accounting for quality change was the main focus of the contributions to the Griliches (1992) volume, each of which focused on a segment of the growing service sector. The general finding of these studies mimics that of the Boskin Commission Report: the inability to fully account for quality change in continuing products and the introduction of new goods leads to an understatement of output growth and productivity growth. In some sectors, output measurement is notoriously difficult. Griliches (1994) lamented the fact that these “unmeasurable” sectors were growing, which made productivity measurement increasingly difficult. Indeed outputs in these sectors are difficult to define, much less measure. Education, health care, financial services, the judiciary (Peyrache and Zago 2016), tax agencies (Alm and Duncan 2014), the provision of public safety (Carrington, Puthucheary, Rose and Yaisawarng 1997), and municipal solid waste collection and disposal (Pérez-López, Prior, Zafra-Gómez, and Plata-Díaz 2016) are

24 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles prominent examples, most of which occur in the public, or non-market, segment of the economy. In these sectors the definition of the service being provided is unclear, and quality concerns loom large. The Atkinson Report (2005) and Schreyer (2012) explore various measurement options. Zieschang discusses productivity measurement in sectors with hard-to-measure output in Chapter 6. He follows the OECD to define these sectors to include high technology industries, real estate, and services, which in turn include distributive services, financial services, health care, and education. He cites three central elements of these sectors: difficulties defining the characteristics associated with outputs, a scarcity of information on the production and accumulation of intellectual property assets, and a lack of sufficiently frequent transactions to permit market valuation. He then uses a capacity utilization function to develop a Fisher-perspective quality-adjusted MFP index that is conditioned on these elements, and he discusses some properties of this index. He then develops a translog approximation, which is exact under certain conditions. He also discusses issues raised by new and disappearing products, and changes in the scope of output, intermediate input, and primary services. Diewert discusses productivity measurement in the public sector in Chapter 7, with a micro orientation toward individual service providers. He analyzes three scenarios: (i) neither output quantities nor output prices are available; (ii) output quantity information is available but output price information is not; and (iii) both output quantity and output price information are available. In the first scenario he proposes to set output growth equal to input growth, in which case productivity growth is zero by construction. In the second scenario he proposes to value outputs either at their unit costs of production, which confronts a difficult cost allocation challenge, or at purchasers’ valuations, which he interprets as quality adjustment factors. Both output valuation options allow for non-zero productivity change. In the rare third scenario conventional productivity measurement techniques are applicable. He devotes most of his attention to the first and second scenarios, in which conventional techniques cannot be applied, and ingenuity is required. In this context he places particular emphasis on the challenges confronting productivity measurement in the education, health care, infrastructure, distribution, and public transportation sectors. In Chapter 17 Lefebvre, Perelman, and Pestieau adopt both micro and macro orientations toward public-sector performance assessment. At the micro level they contrast output, input, and exogenous data that are available for productivity measurement with “ideal” data in rail transport, waste collection, secondary education, and health care. Unsurprisingly, they find available data to be deficient, particularly for their lack of quality information and their failure to incorporate institutional features. At the macro level, they discuss the evaluation of the performance of 28 European welfare states and their evolution through time. The welfare state is a subset of the public sector, with its performance evaluated by indicators of poverty, inequality, unemployment, early school leavers, and life expectancy. Performance is defined as a function of these five indicators, without regard for welfare spending, so that performance coincides with a quantity index of outcomes only. The authors construct three performance indices, one based

Overview of Productivity Analysis 25 on simple unweighted aggregation of scaled indicators used in the original Human Development Report, and a pair of “benefit of the doubt” indices obtained from variants of DEA, in which countries are evaluated on the basis of their ability to maximize the provision of the five welfare-enhancing indicators. They use these indices to estimate change in welfare outcome through time, which grows slightly faster prior to the GFC than after it, and which they use to allocate performance change to improvements in best practice and to catching up, respectively, an exercise related to the “distance to frontier” literature we discuss in subsection 1.4.2. They continue by conducting a test of the cross-country convergence hypothesis, with convergence referring to welfare outcome performance rather than productivity performance, as Inklaar discusses in Chapter 22. They reject the welfare outcome convergence hypothesis. When public sector output quantities are available but output prices are not, one strategy suggested by the Atkinson Report, and reiterated by Diewert in Chapter 7, is to use unit costs of production as proxies for output prices. Grifell-Tatjé and Lovell (2008) study productivity and financial performance at the US Postal Service (USPS) over the period 1972–2004. The USPS reports revenue and an output quantity index, and cost and input quantity indices, which Grifell-Tatjé and Lovell use to construct implicit output and input price indices; as there is a single output quantity index, the cost allocation problem is avoided. They also estimate an efficient unit cost index as an alternative to the implicit output price index to test if the implicit output price index reflects (diminishing) monopoly power. They find the difference between the two proxies to have been small and statistically insignificant over the period 1972–2004. There is another scenario in which unit costs have been used to weight outputs, namely in the private sector when output prices are available but are thought to be distorted, by market power or cross-subsidization, for example. Caves, Christensen, and Swanson (1980) used the neoclassical growth accounting framework to contrast two estimates of US railroad productivity growth during 1951–1974: (i) the conventional approach based on growth in outputs weighted by observed revenue shares that reflect cross-subsidization of passenger service by freight service; and (ii) growth in outputs weighted by estimated cost shares that reflect the structure of production technology. They found, exactly as economic theory predicts, that the replacement of observed revenue share weights with estimated cost share weights reduced the estimated rate of productivity growth from 3.6% per annum to 1.5% per annum. Ironically, a decade after Griliches (1994) lamented the growing “unmeasurable” sector and the publication of the Boskin Commission Report (1996) on difficulties in output measurement, Corrado, Haltiwanger, and Sichel (2005) argued that the fraction of capital that is difficult to measure, accounted for by intangible capital rather than physical capital, which itself is difficult enough to measure, also was growing through time, and also makes accurate productivity measurement increasingly difficult. Measurement difficulties have led the national accounts to treat expenditure on most components of intangible capital as an intermediate expense rather than as investment, which Corrado, Haltiwanger, and Sichel note has potentially serious implications for empirical analyses of business performance and the sources of aggregate economic growth.

26 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles Corrado, Hulten, and Sichel (2005) identify three categories of intangible capital: business investment in computerized information (computer software), innovative property (scientific and non-scientific R&D), and economic competencies (brand equity and firm-specific resources such as organizational capital). Well-established complementarities among various types of intangible capital, and between them and various types of labor skills, mean that organizational capital and skilled labor tend to be bundled in successful businesses, making productivity measurement even more challenging. Nonetheless Corrado, Hulten, and Sichel (2009) accept the challenge by constructing time series data for intangible capital and its three categories, and estimating their contributions to labor productivity growth in the US nonfarm business sector. They find intangible capital deepening to have accounted for about one-quarter of the 1.63% annual labor productivity growth during 1973–1995, and to have accounted for the same share of the much more rapid 3.09% annual labor productivity growth during the subsequent 1995–2003 period. Corrado and Hulten (2014) find similar results for US private industry over the longer 1980–2011 period, with intangible capital deepening accounting for 27% of labor productivity growth. Niebel, O’Mahoney, and Saam (2017) adopt a similar approach to estimating the contribution of “new intangibles” to sectoral labor productivity growth in the European Union during 1995–2007. Their new intangible assets are distinct from ICT assets, and include scientific R&D, firm-specific human capital, and expenditure on market research and advertising, among other components. Their data exhibit considerable variation in the contribution of new intangible assets to labor productivity growth across sectors and countries. Their econometric growth accounting exercise generates statistically significant elasticities of new intangibles on the order of 0.12–0.18, with these estimated elasticities exceeding their factor shares, suggesting the potential for productivity-enhancing resource reallocation.

1.4. Productivity Dispersion Productivity dispersion originates at the individual firm, or even plant, level. Its analysis can be traced back to Schultz (1964, 1975), who studied the ability to rectify disequilibria, departures from satisfaction of first-order optimization conditions. A popular measure of dispersion is a “productivity gap,” expressed as a ratio or a difference, between best and worst performance, or between 90th and 10th deciles, or the interquartile range. Another popular measure is the second moment of the productivity distribution. Productivity dispersion is important because it constrains aggregate productivity; in a widely cited illustration, Hsieh and Klenow (2009) calculate that if productivity dispersion in China and India were reduced to that found in the United States, aggregate productivity would increase by 30%–50% in China and by 40%–60% in India. Much of the literature is devoted to the identification of the sources of productivity dispersion and the promulgation of policies aimed at its reduction.

Overview of Productivity Analysis 27 In Chapter 18 of the Handbook, Bartelsman and Wolf survey alternative measures of productivity dispersion and discuss statistical and economic issues involved in measuring it. They distinguish the preferred physical MFP measure (TFPQ) from the more common revenue factor productivity measure (TFPR) imposed by the data constraint, and they discuss sources and consequences of divergence between the two measures. The problem is that TFPQ (when establishment-level output prices are observed) is an empirical productivity index, but TFPR (when establishment-level output prices are unobserved, and are replaced by an industry output price index common to all producers) is not an empirical index number but often is the only available measure. Revenue productivity is a value rather than a physical concept, which can lead to erroneous inferences about productivity and its dispersion if individual producer prices vary. They also discuss a range of statistical issues that arise in estimating productivity and its dispersion, including endogeneity of input choice and how to deal with it, the use of cost elasticities in growth accounting methods when first-order conditions are violated, the use of stochastic frontier techniques, and how to reduce sensitivity to ubiquitous measurement error. They also provide new evidence of productivity dispersion derived from US and European data. They find similarly large interquartile ranges in the United States and Europe, and less dispersion in gross-output productivity measures than in value- added productivity measures. They attribute much of the observed productivity dispersion to country and industry fixed effects. The productivity dispersion literature is not alone in searching for institutional and other factors behind country, industry, and time fixed effects. The entire productivity literature is, typically of necessity, inundated with unobserved heterogeneity controls, and consequently so is this chapter. Their replacement with variables reflecting the institutional and other sources of these effects would add considerable insight to empirical studies of productivity and its dispersion. The distinction between TFPQ and TFPR arises frequently in the large sample segment of the productivity dispersion literature. Collard-Wexler and De Loecker (2015) show that dispersion and reallocation findings can be extremely sensitive to whether or not one controls for establishment-level price variation, which they suggest would signal variation in mark-ups reflecting market power rather than variation in productivity. They, and Andrews, Criscuolo, and Gal (2016), employ a mark-up correction developed by De Loecker and Warzynski (2012) to convert estimates of TFPR to estimates of TFPQ. Andrews, Criscuolo, and Gal (2016) find differential mark-ups to have accounted for a small portion of the rising productivity gap between frontier and laggard firms in the OECD since 2000. A similar consideration (and an analogous correction) arises on the input side, although the input side remains less frequently studied. De Loecker, Goldberg, Khandelwal, and Pavcnik (2016) develop an analogous framework for correcting for input price variation in a context in which trade reform influences mark-ups through both output and input tariff reductions. Bartelsman and Wolf refer to the issue in Chapter 18.15 Price variation need not reflect variation in market power. Another possibility, frequently discussed in the management literature, is that price variation may reflect variation in the willingness to pay of consumers for customized products, which in turn

28 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles reflects the business strategy of the seller(s). Niche markets are common, with mobile telephones a prominent recent example highlighted in The Economist (2016c). The willingness-to-pay approach to business strategy was introduced by Brandenburger and Stuart (1996), with the difference between consumers’ willingness to pay and sellers’ opportunity cost providing a measure of the value created by businesses and its distribution, a subject we discuss in subsection 1.2.3. From this broad perspective based on business strategy, it is not surprising to find evidence of productivity dispersion. Business seeks profit-driven survival, to which end alternative business strategies generate varying productivities. It would be useful to confine an investigation into productivity dispersion to businesses following similar strategies, since then observed productivity dispersion would reflect varying success in implementing similar strategies, and would signal varying financial performance. This is essentially the approach followed in the productivity convergence literature, with convergence being either unconditional or conditioned on country-specific variables, as exemplified by Rodrick (2013). In this case, conditioning would be on variables characterizing alternative business strategies. However it is measured and whatever its sources, evidence of productivity dispersion has accumulated for over a century, leading Syverson (2011) to characterize inter- plant and inter-firm productivity gaps as “ubiquitous, large and persistent.” We survey this evidence in subsection 1.4.1. In subsection 1.4.2 we explore productivity dynamics, the intertemporal behavior of productivity dispersion. In some circumstances, market forces lead to a reallocation of resources that reduce productivity dispersion, while in others dispersion can be long-lasting, or persistent, and even increase.

1.4.1. Evidence The BLS began publishing its Monthly Labor Review in 1915. Early issues contained numerous empirical studies of (usually labor) productivity dispersion at the plant and company levels. In one study covering 11 sawmills and five production processes, Squires (1917) found inter-plant labor productivity gaps within narrowly defined processes (e.g., tree felling and log-making) in excess of 5 to 1, and unit labor cost gaps ranging from 4 to 1 to 12 to 1. Stewart (1922, 3), US commissioner of labor statistics at the time, summarized numerous similar inter-plant studies across a range of industries and reported comparable dispersion in labor productivity. He concluded by pondering, reasonably, “One asks how a mine that gets but 30.1 pounds per man per day can exist as against a mine securing 371 pounds per day, but with this economic problem we have nothing to do at this time.” Much of the current research on productivity dispersion is aimed at investigating precisely this economic problem. Summarizing several studies across a wide range of industries dating from the 1930s and 1940s, Salter (1960) found similarly large variation in labor productivity, part of which he attributed to “delay in the utilisation of new techniques,” as measured by variation in plant construction date. This vintage effect is a component of the quality of the

Overview of Productivity Analysis 29 capital input we mention in subsection 1.3.3, but it is difficult to quantify and its variation is frequently, and unfortunately, missing from lists of potential sources of productivity dispersion. For a decade during the 1950s and 1960s, the European Productivity Agency published Productivity Measurement Review. The Review contained numerous studies of inter-plant and inter-firm comparisons, usually of labor productivity or its reciprocal, with substantial productivity dispersion being the norm and high/low gaps of 5 to 1 not infrequent. Some studies reported the impact of productivity dispersion on unit labor cost or unit cost, and occasionally on operating ratios such as ROA, demonstrating once again the impact of productivity on business financial performance. Recent evidence on productivity dispersion comes in two complementary forms: large sample evidence, popular in economics, and focused sample evidence, popular in industrial relations and human resource management. The latter approach, also known as insider econometrics, has three steps: (i) interview managers, workers, and others in a firm or firms; (ii) gather relevant data; and (iii) conduct an econometric investigation to test hypotheses of the factors that generate behavior reflected in the data. The focused sample approach complements the large sample approach in two ways: (i) it contains precise measures of dependent and independent variables of interest, including management practices, and (ii) it contains detailed controls for sources of heterogeneity, many of which are unobserved in large sample studies. Shaw (2009) provides a valuable introduction to the focused sample approach, and Ichniowski and Shaw (2012) provide a comprehensive overview. Productivity dispersion has been documented in many large sample studies, including but not limited to Foster, Haltiwanger, and Krizan (2001) [US manufacturing plants in Census of Manufactures years 1977, 1982, 1987 and 1992; US automotive repair shops, 1987–1992]; Eslava, Haltiwanger, Kugler, and Kugler (2004, 2010) [Colombian manufacturing plants, 1982– 1998]; Foster, Haltiwanger, and Syverson (2008) [US manufacturing plants in five Census of Manufactures years]; Bartelsman, Haltiwanger, and Scarpetta (2009, 2013) [establishment-level data across several countries and varying time periods]; Midrigan and Xu (2014) [South Korean manufacturing establishments, 1991–1999]; Asker, Collard-Wexler, and De Loecker (2014) [nine firm- level data sets spanning 40 countries and varying time periods]; and Collard-Wexler and De Loecker (2015) [US steel plants, 1963–2002]. The motivations and issues considered vary widely across studies, but the unanimous finding is one of substantial productivity dispersion whenever and wherever one looks. Productivity dispersion has been documented in many focused sample studies as well, including Ichniowski, Shaw, and Prennushi (1997) [complementarities among human resource practices at steel production lines]; Lazear (2000) [piece rate pay and hourly pay at a firm that installs windshields in cars]; Bartel, Ichniowski, and Shaw (2004) [complementarities among ICT and HRM practices at US valve-making plants]; and Bartel, Ichniowski, Shaw, and Correa (2009) [more complementarities at US and UK valve-making plants]. The last two studies are illustrative of the focused sample approach and its emphasis on the use of specific ICT and HRM practices. Valve-making

30 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles consists of three sequential processes—setup time, run time, and inspection time—and requires three types of computer-based technology and three types of HR practices. Findings indicate that (i) some technologies significantly raise productivity (reduced time) in some processes and not in others; (ii) some HR practices significantly raise productivity in some processes and not in others; (iii) complementarities exist between some IT and HR indicator pairs; and (iv) increases in the use of IT increases the demand for skilled labor (computer skills, programming skills, and engineering skills). In sharp, and useful, contrast to the first finding, Bartel, Ichniowski and Shaw (2004) show that a conventional econometric productivity analysis of the same US valve-making plants based on Census of Manufactures data shows aggregate capital to have no effect on aggregate output (value of shipments less change in inventories), thereby illustrating a virtue of focused-sample studies. We return to focused sample studies in Section 1.5.2, where we consider HRM practices as productivity drivers. In Chapter 10 Garcia-Castro, Ricart, Lieberman, and Balasubramanian provide a somewhat different approach to a focused sample study in which they develop a value- creation and value-capture business model and apply it to the disruptive low-cost no- frills business model of Southwest Airlines, a popular subject of management research, perhaps second only to Toyota.16 Their model, based on Lieberman, Garcia-Castro, and Balasubramanian (2017), is structurally similar to that of Grifell-Tatjé and Lovell in Chapter 9, although terminology differs and content differs in one important way. Their value capture, or price, effect reflects changes in buyers’ willingness to pay and suppliers’ opportunity costs. Their value creation, or quantity, effect consists of an innovation, or productivity, effect, but no replication, or margin, effect, because they expense profit; see our discussion in subsection 1.2.1. Thus the contribution of replication, which plays such an important role at Southwest Airlines through expansion achieved by adding new routes, occurs outside rather than within their analytical framework. Although replication does not increase productivity at Southwest Airlines, it can lead to an increase in industry productivity if it occurs at the expense of less productive carriers, which provides a link to the reallocation literature we discuss in subsection 1.4.2. In their evaluation of the performance of Southwest Airlines, they find replication to have been the dominant, albeit declining, source of value creation, and they find it has increased market share relative to that of the legacy carriers, thereby raising industry productivity. Not surprisingly in light of its business model, they also find growth in value created to have been captured primarily by customers initially, and eventually by employees. They conclude by discussing some changes that Southwest Airlines has been making to its “aging” business model. Focused sample studies can be thought of as successors to business histories that were once so popular and so influential. Since businesses generate productivity growth or decline, and create or destroy value, it is unfortunate that productivity analysts spend so little time studying, and learning from, business history. To provide a few examples, classic and modern: (i) Tarbell (1904) chronicled the rise of the Standard Oil Company, which soon thereafter the US Supreme Court found in violation of the Sherman Antitrust Act; (ii) Chandler (1962) chronicled the development of organizational

Overview of Productivity Analysis 31 structures at duPont, General Motors, and the Standard Oil Company (New Jersey) (a creation of the Supreme Court’s 1911 decision) over a century ago; (iii) in a book widely ignored in the new management practices literature that we explore in subsection 1.5.2, Sloan (1964) chronicled the management practices he developed while CEO of General Motors for nearly a quarter of a century; (iv) the studies collected in Temin (1991) catalogue the uses of information to address organizational problems in largely nineteenth-century businesses; (v) Helper and Henderson (2014) trace the decline of General Motors from 1980 to its bankruptcy in 2009 to its deficient productivity and its inflexible management practices, particularly to its management of relational contracts with its suppliers and employees; and (vi) Brea-Solís, Casadesus-Masanell, and Grifell- Tatjé (2015) study the financial performance of Walmart, contending that Walmart’s business model did not change over a 36-year study period, but variation in the way it was implemented under successive CEOs generated variation in productivity and profit.

1.4.2. Productivity Dynamics Productivity dispersion leads, or might be expected to lead, to a reallocation of resources away from less productive units to more productive units, thereby raising aggregate productivity. This is an old expectation, voiced by Stewart (1922) and explored by Denison (1962, 1974). However evidence suggests that some gaps are stubbornly persist ent. The intertemporal nature of this expectation has spawned the phrase “productivity dynamics” to characterize the study of changes in productivity dispersion through time. Warning: productivity dynamics is a different concept from that of dynamic productivity introduced by Färe, Grosskopf, Margaritis, and Weber in Chapter 5. The microeconomic branch of this literature explores firm productivity dynamics that incorporate entry, exit, and reallocation among incumbents. Industry productivity can increase if market shares of more productive incumbent firms increase at the expense of those of less productive incumbent firms. Industry productivity also can increase, even without an increase in the productivity of any incumbent firm, if productivity levels of new entrants exceed those of incumbent firms, or if productivity levels of incumbent firms exceed those of exiting firms. The macroeconomic branch explores what might be called country productivity dynamics, covering a wide range of topics such as the process of catching up, forging ahead and falling behind (Abramovitz 1986); international productivity convergence or divergence (Baumol 1986; Bernard and Jones 1996a) and whether it is unconditional or conditional, depending on institutions and other country-and time-specific factors (Rodrick 2013); the significance of natural resource endowments, distance to the equator, and other features of geography (Hall and Jones 1999); and the rhetorical question of why are we so rich and they so poor? (Landes 1990). Wolff (2013) surveys the macroeconomic literature, with an emphasis on the convergence hypothesis. The productivity convergence literature addresses two recurring issues. The first issue is convergence of productivity to what? One measure is provided by a time series of the

32 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles standard deviation of individual country productivity levels about the mean productivity level of all countries in the sample (σ-convergence, in the jargon). This measure reveals whether country productivities are converging to a common path, which could result from catching up by laggards or falling behind by leaders, or a combination of the two. Another measure is provided by the coefficient of a regression of productivity growth rates on their initial values (β-convergence). This measure reveals whether countries with relatively high initial productivities tend to grow relatively slowly, and conversely. Inklaar and Diewert (2016) propose a third measure of convergence, a time series of the ratio of actual world productivity, defined as a share-weighted average of individual country productivity levels, to frontier productivity, defined as the maximum productivity level over all countries and all years up to the current year (E-convergence). This measure reveals whether country productivities are converging to (catching up with) best practice in the sample; it combines the flavor of Debreu’s (1951) coefficient of resource utilization with that of the sequential Malmquist productivity index of Tulkens and Vanden Eeckaut (1995), Shestalova (2003), and Oh, Oh, and Lee (2017), although frontier productivity has been defined in other ways, as we discuss in the following. The second issue concerns the identification of those sectors of the economy that are driving or constraining convergence, since identification can direct policies toward the right sectors. Empirical findings have varied with countries, sectors, time periods, currency conversion procedures, and the general quality of data. Using the first convergence measure, Inklaar and Timmer (2009) find support for the finding of Bernard and Jones (1996a) of both labor productivity and TFP σ-convergence in market services, but not in manufacturing, among a sample of advanced OECD countries since 1970. Inklaar and Diewert (2016) study productivity convergence across 38 economies from 1995 to 2011, distinguishing between traded goods and nontraded goods sectors. They find evidence of σ-productivity convergence (especially in the traded sector), but in combination with E-divergence, the latter attributed in part to a compositional shift caused by the growth of China and India. As markets worldwide become less regulated, it becomes increasingly possible and timely to establish the presence of an empirical relationship between convergence, or changes in the relative position countries find themselves in relative to the world frontier, and market forces compelling countries and agents therein to economize. The foundation for the theory of dynamic adjustment that can be broadened into convergence or efficiency analysis has been established utilizing an axiomatic approach by Silva and Stefanou (2003), who laid out a set theoretic approach that was then extended in Silva and Stefanou (2007). Elaboration on the foundation for an adjustment-cost framework by switching to the dynamic directional distance function approach allowed Silva, Lansink, and Stefanou (2015) to deal with an even broader characterization of efficiency and productivity notions. Building on the Luenberger-based approach (using the dynamic directional distance function), Stefanou and his colleagues develop the relationship between the primal and dual forms of productivity (Lansink, Stefanou, and Serra, 2015). Econometrically implementable frameworks for the dynamic adjustment model that address asymmetric dynamic adjustment appear in the review by Hamermesh

Overview of Productivity Analysis 33 and Pfann (1996). Specification and estimation of asymmetric adjustment rates for quasi-fixed factors of production, similar in spirit to the Sickles and Streitweiser (1998) model, are found in Chang and Stefanou (1988), while Luh and Stefanou (1991) provide the modeling setup for estimating productivity growth within a dynamic adjustment framework. Previous efforts to estimate productivity growth in a dynamic adjustment model essentially ignored the adjustment/disequilibrium component of the productivity decomposition. Tests of convergence originating in the economic growth literature (Baumol 1986), determine whether or not there is a closing of the gap between inefficient and efficient firms over time. One approach regresses the firms’ average growth rates in technical efficiency on the log of the carriers’ efficiency scores at the beginning of the sample period. A negative coefficient indicates β-convergence. In other words, the higher a firm’s initial level of efficiency, the slower that level should grow. This phenomenon is the result of the public good–nature of technology, which causes spillover effects from leaders to followers as the laggards learn from the innovators and play “catch-up.” One can also utilize a more sophisticated approach involving the Malmquist productivity index procedure. This method, based on the geometric mean of two Malmquist indices, can account for changes in both technical efficiency (catching up) and changes in frontier technology (innovation). In a study of industrialized countries, Färe, Grosskopf, Norris, and Zhang (1994) note that this decomposition allows for a more comprehensive measure of productivity growth convergence since earlier endeavors failed to distinguish between these two components. Badunenko, Henderson, and Zelenyuk apply this framework to macroeconomic data in Chapter 24, and an application to microeconomic data of the sort used by Andrews, Criscuolo, and Gal (2015, 2016) is a logical extension. Whereas our earlier comparisons of the different methods of calculating technical efficiency necessitated an intertemporal production set, the Malmquist index requires the contemporaneous version. Thus, due to rank considerations, only the DEA approach can be used to calculate the index. The existence of a technology gap may present an additional source of growth, but if nations differ in ability to adopt and absorb new knowledge, then country institutional heterogeneity must also be examined. If follower nations exhibit both a technology gap and a low absorption capacity, then technology’s influence on productivity growth will be ambiguous. The importance of technology transfer has been explored previously. For example, Hultberg, Nadiri, Sickles, and Hultberg (1999) show that technology gaps relative to the United States significantly contribute to follower nations’ aggregate productivity growth in the postwar period. It has also been shown that growth is affected by country heterogeneity, which in turn is highly correlated with various institutional variables. Theoretical studies also point to the importance of openness in accelerating the rate of technology transfer or technology adoption (Parente and Prescott 1993). Bernard and Jones (1996a, 1996b) use a model of TFP that includes the productivity differential within a sector from that of the most productive country. Their results are, again, that manufacturing has not contributed significantly to the overall convergence in OECD countries. Cameron, Proudman, and Redding (1999) expand on the Bernard and Jones model to include a term that is comparable to our efficiency term. They

34 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles look carefully at even more disaggregated data in terms of openness and technology transfers, but only consider the relationship between United Kingdom and the United States. Their results are that the technology gap to the United States plays an important role in UK technology advancement. In Chapter 22, Inklaar surveys the convergence literature and conducts three wide- ranging empirical investigations into the σ-convergence hypothesis. In the first he uses the EU KLEMS database to quantify and to examine the sources of trans-Atlantic productivity convergence (actually, divergence). He finds the source to have been in the ICT- producing sector, in which productivity has grown faster in the United States than in the EU-10. In the second investigation he traces the industry origins of changes in productivity dispersion for a broader set of countries. He finds that productivity convergence has been almost entirely driven by convergence in manufacturing. In the third investigation he examines the extent to which trends in productivity dispersion can be explained by several drivers of productivity change, one of which is a measure of distance to frontier. The other explanatory variables are related to R&D expenditure, the role of high-tech labor and capital, foreign direct investment, and market competition. He finds proximity to the frontier to dampen productivity growth and most of the remaining explanatory variables to enhance productivity growth, though not always significantly. However, the impacts of the remaining explanatory variables do not vary depending on distance to the frontier, suggesting that, for this data set, they do not have an impact on convergence. A variant on the identification of the sectors of the economy that drive or constrain convergence is the identification of the geographic regions of the economy that drive or constrain economic activity in general, convergence in particular. Gennaioli, La Porta, Lopez-de-Silanes, and Shliefer (2013, 2014) study over 1,500 subnational regions from a large number of countries over varying time periods, exploring the influence of geography, natural resource endowments, institutions, human capital, and culture on regional productivity and development. They characterize regional convergence as “slow” and “puzzling,” although they find regional convergence faster in richer countries and in countries with better-regulated capital markets and lower trade barriers. Dispersion, whether inter-firm or inter-regional, reduces the aggregate value of whatever is dispersed. Hsieh and Moretti (2015) apply a spatial equilibrium model to data on 220 US metropolitan areas. This enables them to infer rising inter-city productivity dispersion from rising inter-city wage dispersion, which they attribute to increasing housing supply constraints that limit the ability of workers to reallocate to cities with higher wages and, presumably, higher productivity. They estimate that lowering these constraints to those of the median city would lead to a spatial reallocation of labor that would increase US GDP by 9.5%. Nordhaus (2006) provides insight into the potential value added for productivity research of having detailed spatial data on economic activity. He introduces the G-Econ database (http://gecon.yale.edu) to test a variety of hypotheses concerning the impact of geography on economic activity, including the geographic impacts of global warming on output, which he finds to be larger than previously estimated.

Overview of Productivity Analysis 35

1.4.2.1. Reallocation One of many approaches to the analysis of reallocation is based on a productivity change decomposition proposed by Balk (2003), who reviews the approach in Chapter 2. In this approach, aggregate productivity change is decomposed into four sources: entry of new firms, exit of old firms, productivity change in continuing firms, and redistribution of market shares among continuing firms. If entering firms have above- average productivities, or if exiting firms have below-average productivities, or if the productivities of continuing firms increase, or if market shares of continuing firms are redistributed away from less productive firms toward more productive firms, aggregate productivity increases. One definition of the contribution of reallocation is the sum of entry of new firms, exit of old firms, and redistribution of market shares among continuing firms. As Balk notes, competing decompositions exist, but this is Balk’s preferred decomposition, and it nicely characterizes the reallocation mechanism and its potential to enhance aggregate productivity. Among the competing decompositions, perhaps the most popular is that of Olley and Pakes (1996), whose decomposition contains an additional “cross” component that captures the covariance between changes in incumbents’ productivity and changes in their market share, the objective being to determine whether firms experiencing productivity growth (decline) gain (lose) market share. Empirical evidence of the workings of the reallocation mechanism is widespread and, generally speaking, encouraging; as one would expect, reallocation makes a positive contribution to aggregate productivity growth. A brief summary of three studies illustrates the empirical relevance of the reallocation mechanism. The OECD (2001b) reports findings for several European countries over the 1985–1994 decade. In eight countries, aggregate labor productivity growth has been due primarily to labor productivity growth within continuing establishments, with the role of reallocation being small and variable across countries. However, in five countries, aggregate MFP growth has been due primarily to reallocation, driven by both net entry and redistribution. Foster, Haltiwanger, and Krizan (2006) report findings for a large number of establishments in the US retail trade sector for three census years. The data exhibit large and stable interquartile dispersion in labor productivity across establishments. Nearly all labor productivity growth is attributable to net entry, making reallocation the driving force behind aggregate labor productivity growth. The richness of their data allows the authors to decompose entry into entry by new firms and entry by continuing firms (opening additional establishments), and to decompose exit into exit by firms leaving the sector and exit by continuing firms (closing establishments). The main contributors to net entry were continuing firms opening new establishments and exiting firms leaving the sector altogether. Conditional on survival, they find substantial persistence in terms of relative productivity rankings. Foster, Haltiwanger, and Syverson (2008) report findings for a large number of US establishments in seven-digit manufacturing product categories in five census years. The primary drivers of MFP growth have been growth by continuing establishments, followed by net entry, with reallocation explaining between one-fourth and one-third of aggregate productivity growth.

36 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles Two frontier-based approaches to the investigation of productivity dynamics, reallocation, and convergence have emerged, although there appears to be no cross- fertilization between the two literatures. One frontier-based approach uses macroeconomic data, frequently sourced from the Penn World Tables (PWT), and its analytical foundation is provided by one of the two Malmquist productivity indices we discuss in subsection 1.3.2. This approach seems to have originated with Kumar and Russell (2002), who use PWT data on output, capital, and labor from 57 countries over 1965–1990 to estimate a Malmquist productivity index. An assumption of constant returns to scale enables them to estimate labor productivity change and decompose it into the contributions of catching up, technical change, and capital deepening. They attribute most of the aggregate productivity growth to capital deepening, with technical progress and catch-up together accounting for about 20% of the total. They find evidence of convergence to the global frontier, but no evidence of convergence of developing countries to developed countries. They also find a trend in the labor productivity distribution from unimodal to bimodal to be due entirely to the effect of capital deepening. In Chapter 24, Badunenko, Henderson, and Zelenyuk use updated PWT data covering various countries during varying time periods to extend this Malmquist-based global frontier approach to productivity dynamics and convergence. In one extension they follow the new growth theory by incorporating human capital change as an additional driver of labor productivity change. In another they incorporate a financial development change indicator. Throughout they focus on the sources and convergence of labor productivity growth, using statistical methods to test various hypotheses. They also provide a comprehensive literature survey of the new growth theory and of the Malmquist-based global frontier approach to productivity dynamics and convergence. An alternative frontier-based approach uses microeconomic data, is based on a distinction between “frontier” firms and “laggard” firms, and examines trends in productivity gaps between the two categories of firm. In this approach, forging ahead by frontier firms and catching up by laggard firms both raise aggregate productivity, but have the opposite effects on productivity gaps. The framework is enriched by the existence of two frontiers, a national frontier and a global frontier. The approach is reminiscent of the macroeconomic “catching up, forging ahead and falling behind” thesis of Abramovitz (1986). It is also conceptually similar to the Malmquist productivity index approach, but analytically very different. The construction of national and global frontiers is crucial to the approach. Rather than use econometric or mathematical programming techniques to construct frontiers, this literature defines national frontier firms loosely as “best in nation.” Thus Iacovone and Crespi (2010) define national frontier firms as those in the top quartile of the domestic productivity distribution, and the global frontier as an ill-defined envelope of the best national frontiers. Andrews, Criscuolo, and Gal (2015, 2016) define national and global frontiers in both absolute and percent terms, with fixed and variable number of frontier firms, and report little difference in findings. Bartelsman, Dobbeleare, and

Overview of Productivity Analysis 37 Peters (2015) define frontier firms as belonging to the upper quantiles of the relevant productivity distribution. Empirical evidence obtained from the frontier model of reallocation is relatively scant, but encouraging. Bartelsman, Dobbeleare, and Peters (2015) study a large sample of firms in Germany and the Netherlands over 2000–2008 and report findings across a range of industries. A common but not unanimous finding is one of positive complementarity between both investment in human capital and investment in product innovation and proximity to the frontier. Andrews, Criscuolo, and Gal (2015) study firms belonging to two-digit industries for 23 OECD countries over 2001–2009. They find firms at the global productivity frontier to be on average more productive, larger, more export-oriented, and more profitable than nonfrontier firms, productivity being measured as both labor productivity and MFP. Moreover, they find productivity gaps widening through time. This raises the (unanswered) question of why frontier technologies do not diffuse more rapidly, both from global frontier firms to national frontier firms, and from national frontier firms to domestic laggard firms. Among positive findings, gaps between national and global frontiers vary with educational quality, product and labor market regulations, the quality of markets for risk capital, and the extent of R&D collaboration, each of which is subject to public policy influence. Andrews, Criscuolo, and Gal (2016) apply the frontier model to analyze the recent global productivity slowdown, using the OECD-Orbis data base.17 They find increasing productivity among global frontier firms, and increasing divergence between frontier and nonfrontier firms, in both manufacturing and services, even after including various controls. This leads them to attribute the global productivity slowdown to a slowdown in the diffusion process. They note that diffusion has been slowest in sectors where pro- competitive market reforms have been least extensive, and they also consider the roles of adjustment costs of adopting new technologies, rising entry barriers, and declining contestability of markets in slowing the diffusion process. OECD (2015b, 2016a) and Berlingieri, Blanchenay, and Criscuolo (2017) provide details of the distance to the frontier approach and additional findings, and OECD (2015b, 2016a) discusses the role of public policy directed toward raising aggregate productivity growth, in large part by enhancing diffusion. As Andrews, Criscuolo, and Gal (2015) note, this literature is “very small.” Its growth is to be encouraged, particularly if it incorporates some of the analytical advances appearing in the frontier and Malmquist literature. The concept of distance to the frontier has been adopted by the World Bank as an essential component of its evaluation of the performance of economies in its annual Doing Business reports. The 2017 report covers 190 economies, using 41 indicators for 10 topics, each topic representing a dimension of the cost of doing business (e.g., starting a business, getting credit). Economies receive normalized scores for each indicator within a topic, which are then averaged to create 10 topic scores. Topic scores are then averaged to obtain aggregate “ease of doing business” scores. Distance to frontier is measured as the difference between an economy’s score and the best score attained since a given year.

38 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles Finally, economies are ranked according to their distance to frontier scores, for each topic and for the aggregate ease of doing business.18 This unweighted averaging approach to the construction of composite indicators is widespread but controversial. Cherchye, Moesen, Rogge, and Van Puyenbroeck (2007) criticize the approach and propose a DEA-based “benefit of the doubt” alternative that allows countries to attach different weights to indicators. Underlying data are available at www.doingbusiness.org, and enable one to employ frontier techniques to aggregate indicators into topics, and to aggregate topics into an ease of doing business index, and then to use Malmquist productivity indices to construct annual best-practice frontiers and to measure performance variation across economies and performance change through time. Lefebvre, Perelman, and Pestieau use the benefit-of-the-doubt approach in their evaluation of the welfare performance of the public sector in 28 EU economies in Chapter 17.

1.4.2.2. Persistence The evidence cited in the preceding that reallocation raises aggregate productivity is compelling, and provides evidence of market forces at work. However, this evidence also shows that reallocation does not eliminate productivity dispersion, suggesting the presence of barriers to the working of market forces. The primary finding in this regard is that, conditional on survival, substantial productivity dispersion persists through varying lengths of time. Persistence is measured by assigning units to productivity quantiles through time, which requires long panels. If units remain in the same productivity quantile, or perhaps move no further than to an adjacent productivity quantile, persistence of productivity dispersion is inferred. Andrews, Criscuolo, and Gal (2015) attribute persistence in the OECD to slow diffusion of new global frontier technologies to laggard firms. OECD (2016a) cites a particular example: the uptake of cloud computing in OECD countries has been low, particularly among small firms. OECD (2016a) also emphasizes within-country spatial productivity dispersion, and a slowdown in convergence. This combination of productivity- enhancing reallocation and productivity- constraining persistence prompts a search for the sources of persistence. Banerjee and Moll (2010) mention four potential sources: (i) finance constraints, either inadequate access to financing or adequate access at exorbitant interest rates; (ii) firms of varying productivity self-select into the formal (and taxed) and informal sectors; (iii) political connections enable low-productivity firms to survive; and (iv) a regulatory environment that discriminates against large firms. We discuss some of these sources in subsection 1.5.2. Gibbons and Henderson (2012a, 2012b) suggest another reason why best practices do not diffuse more readily. Their explanation is that many management practices are not based on formal contracts, but on relational contracts involving “. . . a shared understanding of each party’s role in and rewards from achieving cooperation.” These contracts are hard to build and change, causing slow diffusion and persistent productivity gaps. They illustrate with three examples, the “fair” bonus system at

Overview of Productivity Analysis 39 Lincoln Electric, the employee-operated andon cord on the assembly line at Toyota, and the pro-publication philosophy at Merck. Helper and Henderson (2014) apply the relational contract concept to the search for explanations for the decline of General Motors. Yet another possible source of persistence is that business models vary across firms, with some pursuing a financial objective by emphasizing productivity, and others pursuing the same or a different financial objective by pursuing replication, a strategy emphasized by Garcia-Castro, Ricart, Lieberman, and Balasubramanian in Chapter 10. Variation in the quality of management practices, which we discuss in subsection 1.5.2, is another likely source.

1.5. Productivity Drivers Productivity varies across plants and firms and through time, but not randomly. Productivity dispersion has sources, or drivers. It is important to identify drivers in order to adopt business strategies and public policies designed to reduce dispersion and increase aggregate productivity. Two quite different approaches to the identification of drivers of productivity dispersion have developed in the literature, and have been used in two quite different contexts.

1.5.1. Technology-Based Drivers One approach to the identification of the forces driving productivity change is technology-based, specifies two or more drivers, and is typically, although not necessarily, applied to the estimation and decomposition of productivity change through time. In this approach, productivity change is decomposed into technical change (a shift in the best-practice frontier, perhaps caused by the introduction of a new form of ICT) and efficiency change (a movement toward or away from the best-practice frontier, perhaps caused by the diffusion of the new form of ICT). The concept of a best-practice frontier is an essential component of the technology-based approach, but this frontier is typically unobserved, and must be estimated. The search for technology-based drivers thus fits into the “Estimation” part of subsection 1.3.2. If best-practice technology is represented by a production frontier, efficiency change is technical efficiency change. Data requirements are relatively undemanding: output quantities and input quantities for a panel of production units. Analysis is based on distance functions that provide the basis for either of two Malmquist (1953) productivity indices that Russell analyzes in Chapter 4. The Malmquist productivity index proposed by Caves, Christensen, and Diewert (1982) is more popular than the Malmquist productivity index anticipated by Hicks (1961) and Moorsteen (1961), and subsequently proposed by Diewert (1992) and Bjurek (1996), although the latter index

40 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles has more desirable properties. O’Donnell (2012) demonstrates that the Malmquist productivity index proposed by Diewert and Bjurek is a “multiplicatively complete” productivity index because it can be expressed as the ratio of a well-behaved output quantity index to a well-behaved input quantity index, with good behavior requiring the indices to be non-negative, nondecreasing and homogeneous of degree one. He also shows that a multiplicatively complete productivity index can be exhaustively decomposed into the product of three drivers. He illustrates by decomposing the Diewert-Bjurek productivity index into the product of (i) technical change that measures a shift in the production frontier; (ii) technical efficiency change that measures movements toward or away from the production frontier; and (iii) scale-mix efficiency change that measures movements along the production frontier associated with changes in the input mix, the output mix, and the scale of production. However, since the Malmquist productivity index proposed by Caves, Christensen, and Diewert cannot be expressed as the ratio of an output quantity index to an input quantity index without imposing severe restrictions on the structure of technology, it is not generally multiplicatively complete and cannot be exhaustively decomposed into the product of three drivers. If best-practice technology is represented by a cost (or revenue) frontier, efficiency change is cost (or revenue) efficiency change, and it is complemented by technical change measuring a shift in the cost (or revenue) frontier and scale-mix change that measures movements along the cost (or revenue) frontier. Data requirements are somewhat more demanding: output quantities, input prices, and expenditure on inputs for estimation of a cost frontier, and input quantities, output prices and revenue from outputs for estimation of a revenue frontier, as Russell shows in Chapter 4. Grifell-Tatjé and Lovell (2013) propose several analyses of business performance based on the concept of a cost frontier, one of which generalizes cost variance analysis in management accounting. Additional options are available, corresponding to alternative specifications of objectives of and constraints facing firms. In one appealing specification, management is given a budget and told to spend it wisely, a specification that can be traced back to Shephard (1974) and thought of as combining primal and dual approaches. Wise expenditure might be directed toward maximizing output or revenue or ROA. In this case, an output quantity index is based on changes in output quantities as usual, and use of revenue or ROA requires constructing an implicit output price index as well. However, an indirect input quantity index is based on changes in budget-deflated input prices, since input quantities are endogenous choice variables constrained by the exogenous budget and input prices. An indirect productivity index is the ratio of an output quantity index to an indirect input quantity index. The analytical details of (cost or revenue) indirect productivity measurement are available in Färe and Grosskopf (1994). Potential applications are numerous, particularly in the provision of public services such as education, in which agencies receive operating budgets and are expected to maximize outputs such as educational outcomes; Grosskopf, Hayes, Taylor, and Weber (1997) provide an application to public schools. Johnson (1975, 1978)

Overview of Productivity Analysis 41 recounts a private sector example, in which managements at duPont and General Motors allocated funds across product lines with an objective of maximizing the return on these funds. In Chapter 19, Diewert and Fox combine the analytical framework of Shephard (1974) with the analytical framework of Lawrence, Diewert, and Fox (2006), which we mention in section 1.2, to develop a macroeconomic decomposition of value-added growth into an extended set of drivers. They define a cost-constrained value-added function as the maximum value added that can be obtained from a flexible primary input budget. They then develop an analytical decomposition of growth in cost-constrained value added into the product of six drivers: (i) growth in cost-constrained value added efficiency; (ii) growth in net output prices; (iii) growth in primary input quantities; (iv) growth in primary input prices; (v) technical change; and (vi) scale economies. They illustrate their decomposition with data from the corporate and non-corporate nonfinancial sectors of the US economy over the period 1960–2014. Among their findings is a decline in value-added efficiency during recessionary periods when output declines but quasi-fixed inputs cannot be adjusted optimally. They conclude by considering a pair of procedure for aggregating the two decompositions to the entire nonfinancial sector, a top-down approach and a bottom-up approach, which they implement.

1.5.2. Organizational and Institutional Drivers An alternative approach to the identification of the factors driving productivity change is based on the productivity dispersion analysis in section 1.4, and is typically, although not necessarily, applied to an investigation of the sources of productivity variation across producers. Drivers are sorted into two types: organizational factors that originate within the firm and are in principle under management control, and institutional or structural features that are external to the firm and presumably are beyond management control but subject to public policy. We review a few of the more prominent drivers of each type, with an acknowledgement that many of the organizational drivers are strongly correlated, and all could be labeled management practices.

1.5.2.1. Organizational Drivers 1.5.2.1.1. Management Practices The distinguished management consultant Peter Drucker (1954, 71) asserted that “. . . the only thing that differentiates one business from another in any given field is the quality of its management on all levels. And the only way to measure this crucial factor is through a measurement of productivity that shows how well resources are utilized and how much they yield.” In a global research agenda stretching over the past decade, Bloom, Van Reenen, and colleagues have developed the World Management Survey (WMS).19 The survey currently contains data on management practices from over 11,000 firms, primarily in manufacturing but also in retail trade, health care, and education, in 34 countries

42 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles through four different survey waves from 2004 to 2014. Data include quality indices (normalized and ranked from one for worst practice to five for best practice) for 18 management practices in three categories: monitoring, target setting, and people management. Aggregating these quality indices provides an empirical measure of Drucker’s management quality (although recall our reference to benefit of the doubt weighting in subsection 1.4.2). It also enhances the likelihood of discovering the previously missing input in productivity studies noted by Hoch (1955), who called it “entrepreneurial capacity,” and by Mundlak (1961) and Massell (1967), who called its omission “management bias.” More significantly, the WMS data enable one to test hypotheses concerning the drivers of management quality and, in turn, the impact of management quality on various indicators of firm performance. It is dangerous to summarize what is by now a large and rapidly growing body of work, but a few findings are common to the vast majority of studies: (i) there is large cross-country variation, and even larger within-country dispersion, in the quality of management practices; (ii) product market competition is an important driver of the quality of management practices; (iii) ownership matters, with the quality of management practices higher in the private sector than in the public sector, higher in multinational firms than in domestic firms, and higher in professionally managed family firms than in family-managed family firms, particularly primogeniture family firms; (iv) the quality of management practices is positively associated with a range of measures of firm performance, including productivity, profitability as measured by return on capital employed, sales and sales growth, market value as measured by Tobin’s Q, and the probability of survival; and (v) at least a quarter of country productivity gaps with the United States are accounted for by gaps in the quality of management practices. Among the more recent studies based on the WMS, each of which provides references to earlier studies, are Bloom, Genakos, Sadun, and Van Reenen (2012), Bloom, Lemos, Sadun, Scur, and Van Reenen (2014), and Bloom, Sadun and Van Reenen (2016). The WMS contains limited information on American firms. However, Bloom, Brynjolfsson, Foster, Jarmin, Patnaik, Saporta-Eksten, and Van Reenen (2014) report findings based on a recent Management and Organizational Practices Survey (MOPS) of over 30,000 US manufacturing establishments conducted by the US Census Bureau. Their findings complement those based on the WMS, and include (i) enormous dispersion in the quality of management practices; (ii) high correlations between the quality of management practices and firm size, firm location, firm export status, firm employee education, and the intensity of ICT use; and (iii) a high correlation between the quality of management practices and firm performance as measured by productivity (labor productivity and MFP), profitability (operating profit divided by sales), employment growth, and innovation (R&D and patent intensity). In Chapter 12, Benner expresses reservations about the impact of management practices on innovation, suggesting that they may promote relatively minor process innovation at the expense of potentially major product innovation.

Overview of Productivity Analysis 43 The evidence from both surveys is compelling: management matters, for productivity, for financial performance, and for survival. The evidence also identifies several drivers of the quality of management practices, some such as ICT adoption and the use of incentives being organizational in nature, and others such as product market competition and the regulatory environment being institutional in nature. We have mentioned earlier the old notion of management as the missing input. This may have prompted Bloom, Lemos, Sadun, Scur, and Van Reenen (2014) and Bloom, Sadun, and Van Reenen (2016) to consider two alternative views of management, one with management as design from organizational economics, and the other with management as intangible capital, as an input in production technology. With their preferred view of management as intangible capital, they write Y = F(A, K, L, M), in which M is management. Treating intangible capital as a separate input in the production technology is subtly different from treating it as a part of K (Corrado, Hulten, and Sichel 2005) or as a part of L (Bryan 2007). However, the once missing input is no longer missing, and although the interpretation of M is clear, it remains a composite indicator subject to the same concerns we express in subsection 1.4.2 in relation to the construction of distance to the frontier indicators used by the OECD and the World Bank.

1.5.2.1.2. Human Resource Management Practices Human resource management (HRM) practices vary across firms and countries, as does productivity, which prompts a search for a relationship, and perhaps causality. HRM practices are similar to the people management component of management practices, but we treat the topic separately for three reasons: (i) an independent literature exists, (ii) much of its empirical content consists of focused sample studies rather than large sample studies found in the management practices literature; and (iii) large sample studies rarely isolate the impact of the people management component of management practices on productivity. In an early focused sample study, Ichniowski, Shaw, and Prennushi (1997) study 36 homogeneous steel production lines at 17 companies over several months. In their panel, both HRM practices and productivity vary widely. Consistent with economic theory (e.g., Milgrom and Roberts 1990, 1995), they find that clusters of innovative work practices, including incentive pay, teams, flexible job assignments, employment security, and training, have a significant positive effect on productivity, while changes in individual practices have little or no impact on productivity. They also find that clusters of innovative work practices raise product quality. Finally, they find support for two explanations for the failure of best HRM practices to diffuse more widely: slow diffusion of knowledge about the performance of HRM systems, and both pecuniary and nonpecuniary barriers to change at older lines. Lazear (2000) reports the findings of his study of a large auto glass company in which workers install auto windshields, and which gradually changed its compensation from hourly wages to piece-rate pay. Based on a sample of 3,000 workers over a 19-month period, he finds (i) the switch to piece-rate pay led to a 44% gain in output

44 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles per worker; (ii) the company shared the productivity gains with workers in the form of a 10% increase in pay; (iii) the variance in worker productivity increased due to the incentive provided to ambitious workers; and (iv) company earnings increased, although this may have been caused by other factors in addition to the productivity increase. Several focused sample studies find complementarities, not just among HRM practices, but between them and the adoption of various information technologies. Bartel, Ichniowski and Shaw (2004, 2007) find complementarities between HRM practices and computer- based information technologies to increase productivity in terms of reducing setup time, run time, and inspection time at US valve-making plants, and Bartel, Ichniowski, Shaw, and Correra (2009) report similar findings at a larger sample of US and UK valve-making plants. Note the association of productivity with time. In a study of 629 Spanish manufacturing plants, Bayo-Moriones and Huerta-Arribas (2002) attempt to identify determinants of the adoption of production incentives for manual workers. Among the determinants they consider are product market conditions, plant characteristics, work organization, and unions. They find recent increases in product market competition and public ownership to significantly reduce the probability of adoption. They also find the way work is organized, expressed in terms of the number of tasks per job and the share of manual workers in autonomous work teams, and the extent of union influence over workers, to significantly increase the probability of adoption. The extent of plant automation and the magnitude of recent technical changes have no significant impact on the probability of adoption. The authors do not attempt to identify complementarities, and they do not explore the impact of the adoption of production incentives on productivity, but if such incentives tend to raise productivity, then they have uncovered some indirect influences on productivity. In a related study of over 800 Spanish manufacturing plants over six years, Bayo- Moriones, Galdon-Sanchez, and Martinez-de-Morentin (2013) ask whether pay-for- performance practices are likely to be adopted for six occupations, ranging from top executives to sales and production workers. They find sales workers and top executives to be the occupations most likely to adopt such practices, although the nature of pay- for-performance practices varies across occupations. Such practices are less likely to be adopted for production and administrative workers. The idea of adoption of a common pay-for-performance program across all occupations is rejected. Although studies of the impact of HRM practices on productivity are predominantly focused sample studies, a few large sample studies have been conducted. In one such study, which incorporates many attributes of a focused sample study, Black and Lynch (2004) use the Educational Quality of the Workforce National Employers Survey administered by the US Bureau of the Census to construct 1996 cross-section and 1993/ 1996 panel data sets at the individual establishment level. They find that the use of high- performance work practices such as self-managed teams, re-engineering, incentive pay, profit sharing, and employee voice, in conjunction with the adoption of information technologies, including the share of equipment less than four years old and the proportion of non-managers using computers, positively impacts labor productivity. Unlike

Overview of Productivity Analysis 45 many other studies, they are unable to detect any significant complementarities in the cross section and just one in the panel, between unionization and employee voice, as hypothesized by Freeman and Medoff (1984). To the extent that HRM practices can be separated from other management practices, it is expected that good HRM practices enhance producer performance. To the extent that complementarities exist, the impact is magnified. However, there is compelling evidence that many findings are contextual rather than general.

1.5.2.1.3. Adoption of New Technology Even before Salter analyzed vintage effects, Griliches (1957) documented the slow and variable rate of adoption of a new (and currently controversial) agricultural technology, hybrid seed corn, which he attributed to varying profitability of adoption. More generally, David (1990) and Crafts (2004) offer economic historians’ responses to Solow’s “productivity paradox” by providing a broad historical perspective on diffusion lags in the adoption of general-purpose technologies and their consequent delayed impact on productivity. Draca, Sadun, and Van Reenen (2007) summarize what even then was a voluminous literature, breaking it down into macroeconomic studies, industry-level studies, and firm-level studies. Their interpretation of the literature is that it reveals a positive and significant association (but not causality) of ICT with productivity, primarily at the firm level, and largely through complementarities with labor skills and organizational capital, the role of which we discuss in subsection 1.3.3. In Chapter 12, Benner adopts an interdisciplinary approach to propose a managerial resolution to the productivity paradox, a resolution whose roots go back at least to the work of Abernathy (1978). Following Adler, Benner, Brunner, MacDuffie, Osono, Staats, Takeuchi, Tushman, and Winter (2009) and Benner and Tushman (2015), she contends that popular incremental process innovations such as Six Sigma and ISO 9000 convey marginal near-term productivity gains, but at the expense of uncertain but more substantial longer-term productivity gains that might have resulted from successful product innovations. In this view, the opportunity cost of the resources allocated to the adoption of “process management practices” that yield improvements to the existing technology is the possibility of developing an innovative new technology. To the extent that management practices, including HRM practices, can be associated with process innovation, Benner’s analysis bears directly, and critically, on the management practices literature we summarize in the preceding. The pursuit of best management practices appears to raise productivity, but at a potentially high cost; by directing attention away from product innovation, it may preclude the discovery of a radical new product. To paraphrase Gordon (2016), some innovations are more important than others. Benner also provides a useful link to the reallocation literature we survey in subsection 1.4.2 by suggesting that incumbents tend to pursue process innovations, while potential entrants are more likely to pursue product innovations, some successfully. Or, as she puts it so eloquently, “. . . as a firm engages in concerted efforts to produce Blackberries more efficiently, it is actually less likely to create the iPhone. . . .”

46 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles Following up on the product-process innovation distinction, Hall (2011) and Mohnen and Hall (2013) survey the empirical evidence, which suggests that product innovation exerts a statistically significant positive impact on revenue productivity, but process innovation, while it tends to improve business financial performance, has an ambiguous effect on revenue productivity. These incomplete findings have encouraged further research. Product innovation can increase revenue productivity in either of two ways, by increasing output quantities or output prices, and identification is a challenge. Process innovation has an ambiguous effect on revenue productivity, perhaps because it is not intended to raise revenue productivity, but rather is aimed at improving productive efficiency, the impact showing up as an improvement in cost productivity through a reduction in input use. In Chapter 13, Cassiman and Golovko apply the distinction between product innovation and process innovation to explore the complex linkage among the participation of firms in international trade, their innovation activity, and their productivity. They hypothesize that product (but not process) innovation raises productivity and induces firms to self-select into exporting and, eventually, foreign direct investment. They also explore the reverse hypothesis of learning by exporting, which can result from intense competition in foreign markets and from knowledge spillovers from foreign buyers, suppliers, and competitors. This learning can lead to further increases in innovative activity and productivity. As a logical extension, they explore the hypothesis that product innovation and exporting activities are complementary determinants of future productivity growth. After exploring the learning by importing relationship, they explore export–import complementarities on product innovation and productivity. They illustrate these relationships with a large panel of small and medium-sized Spanish manufacturing firms in 20 industries during 1991–2009, and find a “complex dynamic” relating exporting, importing, innovation, and productivity. In a large sample study, Bloom, Brynjolfsson, Foster, Jarmin, Patnaik, Saporta-Eksten, and Van Reenen (2014) relate the use of information technology (IT) to business performance, with management practices providing the intermediate link. Using MOPS, they find three measures of IT usage (IT investment, IT investment per worker, and percent of sales delivered over electronic networks) to be positively correlated with the quality of management practices, which in turn is positively correlated with a variety of business performance indicators. Along similar lines in a pair of focused sample studies Bartel, Ichniowski, and Shaw (2007) and Bartel, Ichniowski, Shaw, and Correa (2009) find strong complementarities between human resource practices and the adoption of new information technologies in enhancing productivity growth. In light of the findings on product versus process innovations, it is worth noting that in each of these studies the technologies being adopted are process, rather than product, technologies. Raymond, Mairesse, Mohnen, and Palm (2015) explore the R&D-to-innovation- to-productivity relationships in a pair of unbalanced panels of Dutch and French manufacturing firms from three waves of the Community Innovation Survey. They find evidence of a lagged positive impact of R&D on innovation, a positive impact of innovation on labor productivity, ambiguous evidence of persistence in innovation, and

Overview of Productivity Analysis 47 strong evidence of persistence in productivity. They also find numerous differences in the relationships between the Netherlands and France. Bos, van Lamoen, and Sanders (2016) also use the Community Innovation Survey, limited to Dutch manufacturing firms, to extend previous work of Mairesse and Mohnen (2002) by specifying and estimating a knowledge production frontier. In their production technology, the innovation output is sales from new or improved products, their innovation inputs consist of a knowledge stock of accumulated innovation expenditures and research labor engaged in R&D activities, and they control for other inputs, cooperation with other institutions, and government funding. They find their innovation inputs to be jointly significant drivers of innovation output, but most of the inter-firm variation in innovation output is unexplained by innovation inputs and controls, and is ascribed to inter-firm variation in innovation efficiency (or innovativeness, or productivity) in the conversion of innovation inputs to innovation output. The finding of innovation inefficiency justifies the extension of the Mairesse and Mohnen knowledge-production function to a knowledge-production frontier, and the finding of innovativeness dispersion is consistent with the widespread finding of productivity dispersion we survey in subsection 1.4.1. Statistics Netherlands (2015) contains a number of firm-level studies exploring various linkages between ICT and productivity. In one study of the impact of ICT capital on sales per worker, ICT capital is disaggregated into eight components. Another distinguishes among product innovation, process innovation, and organizational innovation, and incorporates e-commerce intensity. A third examines the impact of ICT intensity on industry dynamics. Yet another examines the role of ICT in global value chains, distinguishing among ICT-producing firms, ICT-using firms, and non-ICT firms. Findings vary across studies, but a general conclusion is that ICT use enhances productivity, although findings can be sensitive to the definition of productivity, some types of ICT have greater impact than others, and the impact of ICT investment may depend on simultaneous investment in organizational changes. It is apparent that adoption of new ICT investment increases productivity, especially if it is combined with complementary HRM practices. The finding seems subject to two unresolved concerns, however, one involving the distinction between product and process innovation, and the other involving the efficiency with which innovation is adopted and incorporated into business practices.

1.5.2.1.4. Downsizing Global food giant Nestlé SA has embarked on a “Cost and Capital Discipline” program, which it claims has reduced operating cost by CHF1.6 billion in 2014 and 2015 through waste reduction and the leveraging of size and complexity, with an objective of improving return on invested capital.20 Businesses frequently make similar cost-cutting pronouncements, typically in conjunction with quarterly earnings announcements, the objective being to increase competitiveness. It is possible to cut costs in several ways, through the introduction of new resource-saving technology, by reducing waste, by right-sizing, and by recontracting

48 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles with suppliers to gain lower input prices. Only the first two strategies are certain to increase productivity, although the third might. Theory suggests that all four are likely to enhance business financial performance, but the empirical evidence is inconclusive. Attitudes toward cost-cutting vary. The Economist (2016a) complains that cost- cutting announcements rarely are followed by plans to pass the resulting gains on to consumers or employees, thereby exacerbating inequality in the distribution of income. Economists approve of improvements in technology, waste reduction, and right- sizing as components of optimizing behavior, although empirical evidence on their effects, particularly on productivity, is surprisingly mixed. In their study of a sample of US manufacturing plants, Baily, Bartelsman, and Haltiwanger (1996) find productivity gains among plants that increase employment, as well as among plants that reduce employment. In their study of a sample of German firms, Goesaert, Heinz, and Vanormelingen (2015) find that, subsequent to downsizing, both productivity and profitability are largely unchanged, both among downsizing firms responding to a business downturn and, surprisingly, among waste-reducing firms attempting to increase efficiency. In their provocatively titled study of US manufacturing firms, Guthrie and Datta (2008) find downsizing to be negatively correlated with firm profitability, particularly (and unsurprisingly) in growing and R&D-intensive industries. Gandolfi and Hansson (2011) survey the management literature on the causes and consequences of downsizing, and conclude that it has negligible to adverse effects on business financial and organizational outcomes, and largely adverse human impacts, on executioners, victims, and survivors, particularly if downsizing involves shrinking job-training budgets. Perhaps the evidence on the consequences of downsizing is mixed at best because cost-cutting can have adverse unintended consequences within the firm. Fisher and White (2000) adopt a social network view of the firm to show how inefficient cost- cutting can deplete organizational capital, erase organizational memory, and reduce organizational performance, an argument that is easy to support with anecdotal evidence. Schenkel and Teigland (2017) survey the organizational downsizing literature and develop an analytical framework that integrates the concepts of downsizing, social capital, dynamic capabilities, and business performance and competitiveness. Usefully for our purposes, they discuss how ICT capital can be employed to reduce the negative impacts of downsizing.

1.5.2.1.5. Offshore Outsourcing and Global Value Chains Presumably businesses engage in offshore outsourcing to reduce production costs, thereby enhancing their financial performance. If the cost reduction takes the form of an input price reduction, it is likely that offshore outsourcing has no impact on business productivity. Although most of the literature on offshore outsourcing examines its impact on domestic employment, and almost none examines its impact on financial performance, a few studies examine its impact on productivity, and the evidence is mixed. Olsen (2006) surveys the extant literature at aggregate and plant levels, using both labor productivity and MFP, and finds no clear pattern of how the practice affects productivity, with much depending on sector and firm specifics. He does find modest support

Overview of Productivity Analysis 49 for a positive productivity effect, depending on what is outsourced (materials inputs or services inputs, the MS in KLEMS), who is doing the outsourcing (manufacturing or service businesses), and a host of controls for heterogeneity. Amiti and Wei (2009) find a significant positive effect of service offshoring, and a smaller insignificant positive effect of materials offshoring, on both MFP and labor productivity in US manufacturing industries during 1992–2000. Bournakis, Vecchi, and Venturini (2018) examine the productivity impacts of offshore outsourcing in high-tech and low-tech industries in eight OECD countries during 1990–2005. As in previous studies, they find weak support for productivity-enhancing offshore outsourcing, with results varying by industry and whether materials or services are outsourced. Interestingly, they also find some indirect support for Benner’s conjecture in Chapter 12: they find little support for the impact of offshore outsourcing on R&D activities, which they attribute to business’ myopic behav ior, and “. . . which focuses more on short-term cost gains rather than on restructuring and diverting resources towards more innovative activities.” Offshore outsourcing can be an end in itself or, increasingly, it can be a link in a larger global value chain (GVC). To illustrate, an example of a GVC is the iPod, which prior to its discontinuation was designed in the United States, assembled in China by Taiwanese companies using more than 100 components manufactured around the world, with logistics handled in Hong Kong.21 Reijnders, Timmer, and Ye (2016) and Timmer and Ye in Chapter 21 examine alternative characteristics of GVCs. General findings are similar in both advanced and emerging economies, and include (i) increasing fragmentation of production; (ii) a strong bias to technical change that favors capital (particularly ICT capital) and high-skill labor; (iii) increasing specialization; and declining low-skill labor value-added shares. These findings within GVCs are consistent with domestic findings of capital-skill complementarity reinforced by skill-biased technical change we discuss in subsection 1.2.3 in the context of growing inequality. In Chapter 21, Timmer and Ye show how to analyze production, technical change and its bias, factor demand elasticities and their cost shares, and productivity change in GVCs, illustrating their methodology using a KLEMS approach and the World Input- Output Database. They provide two empirical applications, one to the GVC of German automobiles and the other to 240 manufacturing GVCs, both during 1995–2007. In the former application they are able to allocate GVC productivity growth of 0.99% to the German automobile industry itself (0.73%) and to other industries in the GVC (0.26%). In the latter application they find, consistent with non-GVC studies, strong complementarity between capital and high-skill labor. They also attribute growing cost shares of capital and high-skill labor, and declining cost shares of low-and medium-skill labor, primarily to biased technical change, with input price effects being small. This decomposition of changes in cost shares into driving sources is important in its own right, and has widespread potential applicability. Recent evidence suggests that offshoring is a two-way street, with reshoring, the practice of returning production to the home country, becoming more common. De Backer, Menon, Desnoyers-James, and Moussiegt (2016) summarize the economic factors at work and survey the evidence. The factors favoring reshoring include an eroding

50 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles offshore cost advantage, increasing supply risk in longer and more complex GVCs, the need to be close to markets, lagging domestic innovation, and endangered intellectual property. However, they find limited impacts on the home country of reshoring to date, with the major finding being that reshoring leads to a large increase in investment in high-tech capital and a small increase in employment, with most of that being of high- skill labor. Many home country jobs that were offshored are gone forever. Adidas, a German sporting goods firm, provides an excellent recent example. It has outsourced the manufacture of sport shoes, primarily to Asian countries, for many years. However, growing labor shortages and rising labor costs, in conjunction with GVCs that can take up to 18 months from design to delivery, have prompted Adidas to reshore production to Germany. Reshoring will exploit advanced technology (including robotic cutting, 3D printing, and computerized knitting) to shorten the supply chain to less than a week, with a potentially large impact on productivity; once again, productivity is measured in terms of time to completion.22 Not all production fragmentation involves offshoring, with or without reshoring. Using data from the 2007 US Census of Manufactures, Fort (2017) finds a positive relationship between a firm’s use of ICT and its decision to fragment production, either offshore or domestically. She also finds domestic fragmentation to be far more prevalent than offshoring, a finding she attributes to complementarities between ICT and worker skill, which is generally higher in the United States than in countries to which US firms tend to offshore services.

1.5.2.2. Institutional Drivers We know that some business practices enhance productivity, but what institutional features enhance or retard their adoption and diffusion? The work of North (1990), co-recipient of the 1993 Nobel Prize in Economic Sciences “for having renewed research in economic history by applying economic theory and quantitative methods in order to explain economic and institutional change” (The Royal Swedish Academy of Sciences 1993). has inspired research into the impact of policies and institutions, including public infrastructure, on economic performance, including private productivity. Hall and Jones (1999) argue that “. . . the primary, fundamental determinant of a country’s long-run economic performance is its social infrastructure,” consisting of institutions and government policies. The challenge, of course, is to define and construct an index of social infrastructure. They construct such an index, and show that it exerts positive impacts on a variety of indicators of economic performance, including capital accumulation, educational attainment, and productivity, and therefore on per capita income. Easterly and Levine (2001) also emphasize the role of policies, such as legal systems, property rights, infrastructure, regulations, and taxes, that influence both factor accumulation and productivity, and they note that policy differences do not have to be large to matter, since “[s]mall differences can have dramatic long run implications.” More recently, Ègert (2016) examines the effect variation in the quality of institutions, essentially the rule of law and law enforcement, on productivity, using a panel of OECD countries. He shows that higher quality institutions amplify the

Overview of Productivity Analysis 51 productivity-enhancing impact of R&D spending, although why R&D spending should provide the conduit is left unexplained. Hopenhayn (2014) provides an analytical framework for investigating the impact of institutions on economic performance, and surveys some of the more prominent institutions that retard performance. We touch on some of these institutional features in the following.

1.5.2.2.1. Competition Perhaps the most prominent finding to emerge from Bloom and Van Reenen (2007) and their subsequent studies is the importance of product market competition as a determinant of the quality of management practices, and thus of the economic and financial performance of firms. This finding is robust to alternative measures of product market competition, including domestic competition, international competition, and competition as perceived by management. Van Reenen (2011, 306), quoting Adam Smith (“Monopoly . . . is a great enemy to good management”), discusses the common finding that increased product market competition raises aggregate productivity growth, and a less common finding that it does so without substantially reducing productivity dispersion, to which we refer in section 1.4. He also claims that “[p]erhaps the most common form of a competition shock is from trade liberalization” (314). The OECD (2015b, 48) lists three channels through which trade exposure raises productivity: (i) trade openness increases competition, which promotes the productivity-enhancing reallocation we discuss in subsection 1.4.2; (ii) trade and foreign direct investment promote knowledge flows among global suppliers and customers, and within multinational firms, enhancing productivity convergence toward global frontiers; and (iii) trade openness increases effective market sizes, which raises potential productivity gains and profits from adoption of foreign technologies. The empirical literature investigating these channels is voluminous. To cite one recent example, Eslava, Haltiwanger, Kugler, and Kugler (2013) examine the impact of trade liberalization that reduced both the average level and the inter-industry variation in effective tariffs, on manufacturing plant productivity in Colombia during 1982–1998. They find liberalization to have increased exit, raised productivity within continuing plants, and improved resource allocation among continuing plants, all of which combined to raise aggregate productivity. The data constraint prevents them from considering nontariff barriers or the impact of tariff reform on entry. In Chapter 14, De Loecker and Van Biesebroeck survey the recent literature. They consider a range of approaches to estimate the impact of changes in international competition, perhaps but not exclusively through reduced tariff barriers, on productivity. A vast empirical literature suggests that trade liberalization raises aggregate productivity in two ways that overlap with the three channels identified by the OECD (2015b): (i) by raising the minimum level of productivity necessary for survival, and (ii) by reallocating resources toward more productive firms. However, three related themes pervade their discussion. One is yet another warning of the danger in using deflated revenue as an output indicator when measuring productivity, a danger we first encountered in subsection 1.4.1. A second is the difficulty in separating the impacts of increased

52 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles competition on output quantities and output prices, and the development of strategies for decomposing revenue productivity change into a mark-up change component and a productivity change component. They develop three such strategies, borrowed from the theoretical industrial organization and international trade literatures. The third theme is that, while trade liberalization, by enlarging the relevant market, has the potential to increase competition, and hence productivity, whether it actually has this effect is contextual.

1.5.2.2.2. Regulation Regulation can affect productivity in three ways: (i) regulation in an industry can affect productivity in the same industry, (ii) regulation in an industry can affect productivity in other industries using the product of the regulated industry as an input, and (iii) labor market regulation can affect productivity in industries employing regulated labor. In each case, the regulatory impact can take either of two forms: it can lower the mean or increase the dispersion of the productivity distribution. Andrews and Cingano (2014) construct a sample of over 800 country-industry observations across 21 OECD countries in 2005, with an objective of exploring the impact of cross-country variation in product market, labor market, and credit market regulations on cross-country productivity distributions. They find positive relationships between productivity dispersion among existing producers and each of employment protection legislation, product market regulations, and restrictions on foreign direct investment. However, credit market imperfections work differently, lowering the mean of the productivity distribution rather than increasing its dispersion, implying that financial frictions influence entry decisions rather than the productivity distribution of incumbent producers. They conclude that reducing product and labor market entry barriers in each country to the lowest levels observed in the European Union would reduce misallocation by half and increase aggregate labor productivity by 15%. Égert and Wanner (2016) describe the OECD’s suite of indicators of anti-competitive regulation in the economy, REGIMPACT. Components include an economy-wide product market regulation indicator, seven network industry regulation indicators, four professional services regulation indicators, and a retail trade regulation indicator. Each of these indicators varies widely across OECD countries, and all decline through time— two features that make them useful in studies of the impact of cross-country variation in regulation on cross-country productivity distributions. Across a range of regressions, REGIMPACT has a statistically significant negative impact on labor productivity, and a negative impact that is frequently statistically significant on MFP. Each of the following studies uses this database. Bourlès, Cette, Lopez, Mairesse, and Nicoletti (2013) use REGIMPACT to examine the indirect impact of regulation in one industry on the productivity of industries using that industry’s product. Based on a panel of 20 industries in 15 OECD countries over a 24-year period, they find strong evidence that upstream regulation retards downstream productivity. Additionally, using the distance to the frontier concept, they find the adverse impact to be greatest for firms closest to the global frontier, creating a catch-up

Overview of Productivity Analysis 53 effect for laggard firms. Égert (2016) adds two additional potential productivity drivers, innovation intensity and trade openness, to a similar panel. He finds a negative impact of labor market regulations and positive impacts of innovation intensity and trade openness, and complementarity between labor and product market regulations. Cette, Lopez, and Mairesse (2016) study another similar panel, examining the impacts of product and labor market regulations on productivity in downstream industries. They find potential long-term productivity gains on the order of 2.5% and 1.9%, respectively, if all countries adopted “lightest-practice” regulations, defined as the mean of the three lowest regulatory burdens in their sample. In a subsequent study Cette, Lopez, and Mairesse (2017) examine the impact of upstream regulatory burden indicators on downstream MFP in 15 OECD countries, and they find a statistically significant negative impact. They also find a mostly significant negative impact of upstream regulatory burden on downstream investment in ICT and R&D capital. Empirical findings across a wide range of data sets are consistent, with one another and with the predictions of economic theory. Anti-competitive product and labor market regulations, and constraints on access to or the cost of capital, all have adverse consequences for productivity, either direct or indirect, by affecting either the mean or the dispersion of the productivity distribution. Evidence on the negative impact of upstream regulation on downstream productivity is particularly compelling.

1.5.2.2.3. Financial Frictions and Credit Constraints Midrigan and Xu (2014) study financial frictions such as borrowing constraints, arguing that they reduce aggregate productivity through two channels: (i) they distort entry and technology adoption decisions, reducing the productivity of those producers, and (ii) they generate different rates of return to capital across producers, causing misallocation and further reducing productivity. Using establishment data from Korea, with its well-developed financial system, and China and Columbia, both with less-developed financial systems, they find the first channel to be more important than the second, reinforcing the finding of Andrews and Cingano (2014). They attribute the relatively small misallocation effect among incumbents to the ability of financially constrained but nonetheless more productive incumbents to exploit retained earnings as a source of capital. Moll (2014) also emphasizes the ability of accumulated internal funds to moderate capital misallocation caused by financial frictions. Banerjee and Duflo (2014) trace the persistent misallocation of capital among borrowing firms in India to the withdrawal and subsequent reimposition of credit constraints, although their focus is on profits rather than productivity, and they also find the extent of misallocation to be sensitive to the ability to self-finance. Ferrando and Ruggieri (2015) argue that the impact of financial frictions on a firm’s productivity depends on its financial structure. They use the Amadeus database23 to construct a sample of over 5 million firm-year observations across nine sectors in eight Euro-area countries during 1995–2011, a period that includes the GFC. They construct a synthetic indicator of firm financial constraint as a function of financial leverage, debt burden, cash holdings, and firm controls. They find significant negative impacts of their

54 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles financial constraint indicator on labor productivity across most sectors and countries, with the impacts being most severe in “innovative” industries. Fernandes and Ferreira (2017) examine the impact of tightening financing constraints caused by the GFC on firm employment decisions, using Portuguese linked employer–employee data during 2000–2012. They construct a similar indicator of firm financial constraint as a function of external finance dependence, asset tangibility, importance of trade credit, reliance on short-term debt, and a firm size-age index. Firm employment decisions are expressed as the share of workers hired on short-term contracts. Their main finding is that, subsequent to the crisis, firms with above-median financial constraint increased the share of short-term hires in total hires. The authors also suggest that this decision has productivity implications, since relatively constrained firms prefer the flexibility of short- term contracts to the higher productivity associated with permanent contracts. To test this conjecture, they regress labor productivity on the share of short-term workers and a host of firm controls, and they find a statistically negative impact. Caselli and Gennaioli (2013) follow Burkhart, Panunzi, and Shliefer (2003) by linking institutional failures such as financial frictions with ownership. We have noted the finding of Bloom and Van Reenen (2007), repeated in many subsequent studies, that family-owned firms having a family CEO chosen by primogeniture have much lower quality of management practices than other firms, and consequently underperform other firms on a range of economic and financial indicators. Caselli and Gennaioli do not rely on primogeniture, but they do study the intergenerational transfer of management in family firms, a practice they call “dynastic management,” and Firfiray, Larraza- Kintana, and Gómez-Mejia call the “protection of socioemotional wealth” in Chapter 11. They argue that financial frictions hinder the market for corporate control by deterring lending and investment, restricting financing opportunities to both talented outsiders and talented descendants, who would otherwise invest more in the family firm than their less talented siblings. As a result, poorly functioning financial institutions reward less talented descendants, thereby adversely affecting productivity in dynastic family firms, which predominate in developing countries. Financial constraints plague global supply chains, and have motivated the growth of a non-bank “fintech” industry designed to relax these constraints. Nonetheless, The Economist (2017) reports that the vast majority of global supply chains lack an adequate financing program, which, by raising transaction costs, reduces their productivity. Financial constraints also influence modes of production. Using a large panel of US manufacturing plants, Andersen (2017) finds that credit constraints distort the asset mix toward tangible assets that can serve as collateral. This in turn leads to a quantitatively large and statistically significant increase in pollution emissions, an effect that is pronounced in industries that rely on external financing. Unlike other studies, Blancard, Boussemart, Briec, and Kerstens (2006) use the frontier techniques we mention in subsection 1.3.2 to estimate the impact of credit constraints on the financial performance of a sample of French farmers. They find financially unconstrained farmers (whose credit constraint is not binding) to be larger, more efficient, and more successful financially than those having binding credit constraints.

Overview of Productivity Analysis 55 Their use of frontier techniques also enables them to estimate shadow prices of the constraints.

1.5.2.2.4. Costs of Doing Business We have discussed regulation and finance constraints, both of which impose costs that reduce business productivity. There are other costs of doing business, and an associated literature examining the mechanisms through which these costs influence productivity. The World Bank’s Doing Business project was initiated in 2002 and provides objective measures of business regulations and their enforcement. The 2017 edition includes 11 indicators, each with several components, most of which are available for most of 190 economies. The 11 indicators measure the (money and time) costs of starting a business, dealing with construction permits, getting electricity, registering property, getting credit, protecting minority investors, paying taxes, trading across borders, enforcing contracts, resolving insolvency, and labor market regulation. Barseghyan (2008) uses the World Bank’s Cost of Doing Business data set to examine the impact of cross-country variation in entry costs on productivity. He regresses output per worker and TFP on an entry cost indicator, including all official fees that must be paid to complete legal procedures for starting a business, and incorporates a number of institutional controls. He finds that higher entry costs significantly reduce labor productivity, primarily by reducing its MFP component. Moscoso Boedo and Mukoyama (2012) also use the Cost of Doing Business data set, and they add exit costs, consisting of the cost of advance notice requirements, severance payments and penalties due when terminating a worker, to their entry cost indicator, consisting of the monetary and time costs of starting and licensing a business. Both costs and their components vary dramatically across countries, and both reduce productivity. Entry costs lower productivity by keeping low-productivity establishments in business, and exit costs lower productivity by dampening the reallocation of labor from low-productivity to high-productivity establishments. The authors calculate that raising the two costs from their US levels to those of the average low-income countries reduces aggregate GDP by as much as one-third. Industrial policies that keep low-productivity firms in business, and discourage more productive firms from investing, have led to the “zombie firm” phenomenon. Caballero, Hoshi, and Kashyap (2008) study the phenomenon during the Japanese macroeconomic stagnation of the 1990s. They define zombie firms as potentially receiving subsidized bank credit, and they focus on credit misallocation resulting from zombie lending, which they attribute to relationship banking and regulatory forbearance, both of which keep zombie firms from exiting. In a large sample of Japanese firms, they find an increase in the share of zombies in an industry to be associated with a decline in investment and employment growth for non-zombies, and a widening productivity gap between non-zombies and zombies. Adalet McGowan, Andrews, and Millot (2017) define zombie firms as being at least 10 years old with an interest coverage ratio (the ratio of operating income to interest expense) of less than one for the preceding three years. In a large sample of OECD firms, they find (i) zombie firms to have increased

56 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles in number and market share since 2000; (ii) the increasing survival of zombie firms congests markets, restricting exit, constraining the growth of more productive incumbent firms and raising barriers to entry for new firms; (iii) resources devoted to zombie firms reduce productivity-enhancing resource allocation that constrains investment and employment in more productive firms; and (iv) an increased productivity gap between zombie firms and more productive firms; all of which lead to (v) reduced potential output through two channels, business investment and MFP growth. Their primary policy prescription is to reduce credit and other barriers to the exit of zombie firms. Complementarities can be positive, as with combinations of HR practices and types of ICT adoption, but they can be negative as well. Bergoeing, Loayza, and Piguillem (2016) contend that entry barriers and exit barriers are complements; each imposes costs, and they reinforce each other’s negative impact on productivity. They use the Cost of Doing Business data set to estimate the impact of these barriers on the gap between US and developing country per capita output. They find entry and exit barriers to account for roughly half of the gap between the United States and the median developing country, half of which is accounted for by complementarities. The policy implication is that removal, or lowering, of both entry and exit barriers have a greater productivity impact than removing or lowering either of them separately. Entry costs and other costs of doing business have a second depressing effect on productivity. Several writers have documented that high costs of doing business divert business from the formal sector to the informal sector, where productivity is lower than in the formal sector, due in part to the small size and inefficiencies of enterprises in the informal sector. D’Erasmo and Moscoso Boedo (2012) argue that high costs of doing business reduce aggregate MFP by encouraging the growth of the informal sector, where firms tend to be small and less productive than their formal-sector counterparts. This misallocation of capital reduces aggregate MFP; based on World Bank Cost of Doing Business data, they calculate this reduction to be “up to 25%.” La Porta and Shleifer (2014) emphasize the fundamental differences between firms in the formal and informal sectors, arguing that informal firms are long-lived and rarely move to the formal sector, and that informality is reduced only by economic development.

1.5.2.2.5. Home Production Becker’s (1965) analysis of household time allocation was in large part responsible for his receipt of the 1992 Nobel Prize in Economic Sciences for “. . . having extended the domain of microeconomic analysis to a wide range of human behaviour and interaction, including non-market behaviour” (Royal Swedish Academy of Sciences 1992). We know that much non-market economic activity, home production in particular, is not captured by the national accounts. For our purposes, the relevant questions are how to measure the productivity of home production and whether incorporation of home production into the national accounts would have an impact on aggregate productivity, and then on social economic progress, as we discuss in subsection 1.2.3. The OECD (2002, Annex 2) defines household production for own use as comprising “. . . those activities that are carried out by household unincorporated enterprises that

Overview of Productivity Analysis 57 are not involved in market production. By definition, such enterprises are excluded from the informal sector” (which is engaged in market production). Bridgman (2016) summarizes the efforts of the US Bureau of Economic Analysis to construct satellite accounts that estimate the value of household production. These accounts suggest that including home production raises GDP by 37% in 1965 and by 23% in 2014, the decline being attributable to increasing female labor force participation. Poissonnier and Roy (2015) report on the development of a satellite household account for France. Their findings suggest that incorporating household production in the national accounts would increase GDP by a third while reducing its rate of growth, and increase disposable income by half. They also conduct sensitivity analyses of various methodological issues, including the use of gross or net wages, and minimum or living wages, to value household labor. The OECD (2011) provides preliminary estimates of the value of household production of non-market services, with the ultimate objective of comparing material well-being across countries. Among their many findings, they conclude that national estimates are acutely sensitive to the valuation of household labor using replacement cost or opportunity cost methodologies, although international comparisons are not. Schreyer and Diewert (2014) apply Becker’s household time allocation model to determine the conditions under which the replacement cost or opportunity cost approaches are the appropriate way to value household labor, and they develop a cost of living index for Becker’s full income and full consumption (of market goods, work at home, hired labor services, and leisure). This leads them to an international comparison of GDP growth rates with, and without, household production included. They characterize the differences as “not-insignificant.” Their work constitutes an important analytical step toward household productivity measurement, although much more analytical and empirical work is needed.

Notes 1. Blázquez-Gómez and Grifell-Tatjé (2011) find the Spanish regulator to have exhibited a pro-industry, anti-consumer bias during the 1988–1998 period, with the estimated value of the intended fraction having been outside [0, 1] for the majority of electricity distribution companies. 2. Here and henceforth we refer to all aggregate output measures such as GDP and GNP as “output,” except when a precise definition is necessary. 3. https://data.oecd.org/lprdty/unit-labour-costs.htm (accessed October 24, 2016). 4. http://stats.oecd.org/# (accessed October 24, 2016). 5. For more on social economic progress, see Grifell-Tatjé, Lovell, and Turon (2016). 6. http:// w ww.oecdobserver.org/ news/ f ullstory.php/ aid/ 5 548/ T he_ productivity_ and_ equality_nexus.html (accessed October 24, 2016) 7. This example is one of many suggesting potential complementarities between academics and consultancies such as the McKinsey Global Institute. Lewis (2004) provides a readable account of the Institute’s forays into the measurement of productivity and its determinants, at both firm and country levels.

58 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles 8. Alan B. Krueger served as chairman of President Obama’s Council of Economic Advisors from 2011 to 2013. 9. http://www.bls.gov/mfp/data.htm (accessed October 24, 2016). 10. http://www.bls.gov/mfp/mprdload.htm (accessed October 24, 2016). 11. See also www.bls.gov/bls/productivity.htm (accessed October 24, 2016). 12. See http://www.oecd.org/economy (accessed October 24, 2016). 13. See http://www.bls.gov/news.release/pdf/prod3.pdf (accessed October 24, 2016). 14. See also www.bls.gov/ppi/qualityadjustment.pdf (accessed October 24, 2016). 15. Haltiwanger (2016) provides a critical overview of the TFPQ/TFPR literature. 16. A December 30, 2016, Google Scholar search for “the Toyota production system” turned up about 257,000 results. 17. http://w ww.bvdinfo.com/en-gb/our-products/company-information/international- products/orbis (accessed January 4, 2017). 18. This brief summary masks many details, which are available at www.doingbusiness.org/ data/distance-to-frontier (accessed November 22, 2016). 19. http://worldmanagementsurvey.org (accessed November 1, 2016). 20. http://www.nestle.com/asset-library/documents/library/presentations/investors_events/ investor-seminar-2016/nis-2016-14.pdf (accessed January 31, 2017). 21. Timmer, Erumban, Los, Stehrer, and de Vries (2014) attribute this example to Dedrick, Kraemer, and Linden (2010), who also estimate how profit is distributed along the GVC. 22. www.adidas-group.com//en/group/stories-copy/specialty/adidas-future-manufacturing/ 23. http://w ww.bvdinfo.com/en-gb/our-products/company-information/international- products/amadeus (accessed January 4, 2017).

References Abernathy, W. J. 1978. The Productivity Dilemma. Baltimore, MD: Johns Hopkins University Press. Abramovitz, M. 1986. “Catching Up, Forging Ahead, and Falling Behind.” Journal of Economic History 46: 385–406. Adalet McGowan, M., D. Andrews, and V. Millot. 2017. “The Walking Dead? Zombie Firms and Productivity Performance in OECD Countries.” OECD Economics Department Working Papers No. 1372. www.oecd.org/eco/The-Walking-Dead-Zombie-Firms-and-Productivity- Performance-in-OECD-Countries.pdf. Adler, P., M. J. Benner, D. Brunner, P. MacDuffie, E. Osono, B. Staats, H. Takeuchi, M. L. Tushman, and S. G. Winter. 2009. “Perspectives on the Productivity Dilemma.” Journal of Operations Management 27: 99–113. Ahn, S. C., D. Good, and R. C. Sickles. 2000. “Estimation of Long-Run Inefficiency Levels: A Dynamic Frontier Approach.” Econometric Reviews 19: 461–92. Aigner, D., C. A. K. Lovell, and P. Schmidt. 1977. “Formulation and Estimation of Stochastic Frontier Production Function Models.” Journal of Econometrics 6: 21–37. Albrizio, S., T. Kozluk, and V. Zipperer. 2017. “Environmental Policies and Productivity Growth: Evidence across Industries and Firms.” Journal of Environmental Economics and Management 81: 209–226. Alchian, A. 1950. “Uncertainty, Evolution and Economic Theory.” Journal of Political Economy 58: 211–222.

Overview of Productivity Analysis 59 Alm, J., and D. Duncan. 2014. “Estimating Tax Agency Efficiency.” Public Budgeting & Finance 34: 92–110. Alvarez, A., and P. Schmidt. 2006. “Is Skill More Important Than Luck in Explaining Fish Catches?” Journal of Productivity Analysis 26: 15–25. Amiti, M., and S.-J. Wei. 2009. “Service Offshoring and Productivity: Evidence from the US.” The World Economy 33: 203–220. Andersen, D. C. 2017. “Do Credit Constraints Favor Dirty Production? Theory and Plant-Level Evidence.” Journal of Environmental Economics and Management 84: 189–208. Andrews, D., and F. Cingano. 2014. “Public Policy and Resource Allocation: Evidence from Firms in OECD Countries.” Economic Policy 29: 253–296. Andrews, D., C. Criscuolo, and P. N. Gal. 2015. “Frontier Firms, Technology Diffusion and Public Policy: Micro Evidence from OECD Countries.” OECD Productivity Working Papers No. 02. http://www.oecd-ilibrary.org/economics/frontier-firms-technology-diffusion-and- public-policy_5jrql2q2jj7b-en. Andrews, D., C. Criscuolo, and P. N. Gal. 2016. “The Best versus the Rest: The Global Productivity Slowdown, Divergence across Firms and the Role of Public Policy.” OECD Productivity Working Papers No. 05. http://www.oecd-ilibrary.org/economics/the-best- versus-the-rest_63629cc9-en. Arrow, K. J. 1962. “The Economic Implications of Learning by Doing.” Review of Economic Studies 29: 155–173. Asker, J., A. Collard-Wexler, and J. De Loecker. 2014. “Dynamic Inputs and Resource (Mis) Allocation.” Journal of Political Economy 122: 1013–1063. Atkinson, T. 2005. “Measurement of Government Output and Productivity for the National Accounts.” Atkinson Review: Final Report, HMSO. http://web.ons.gov.uk/ons/search/ index.html?newquery=atkinson+report. Autor, D. H. 2015. “Why Are There Still So Many Jobs? The History and Future of Workplace Automation.” Journal of Economic Perspectives 29: 3–30. Ayres, C. E. 1944. The Theory of Economic Progress. Chapel Hill: University of North Carolina Press. Baily, M. N., E. J. Bartelsman, and J. Haltiwanger. 1996. “Downsizing and Productivity Growth: Myth or Reality?” Small Business Economics 8: 259–278. Balk, B. M. 2003. “The Residual: On Monitoring and Benchmarking Firms, Industries and Economies with Respect to Productivity.” Journal of Productivity Analysis 20: 5–48. Balk, B. M. 2008. Price and Quantity Index Numbers: Models for Measuring Aggregate Change and Difference. New York: Cambridge University Press. Balk, B. M. 2009. “On the Relation between Gross Output-and Value Added- Based Productivity Measures: The Importance of the Domar Factor.” Macroeconomic Dynamics 13(Suppl 2): 241–267. Banerjee, A. V., and E. Duflo. 2014. “Do Firms Want to Borrow More? Testing Credit Constraints Using a Directed Lending Program.” Review of Economic Studies 81: 572–607. Banerjee, A. V., and B. Moll. 2010. “Why Does Misallocation Persist?” American Economic Journal: Macroeconomics 2: 189–206. Banks, G. 2015. “Institutions to Promote Pro-Productivity Policies.” OECD Productivity Working Papers No. 01. http://www.oecd-ilibrary.org/economics/institutions-to-promote- pro-productivity-policies_5jrql2tsvh41-en. Barseghyan, L. 2008. “Entry Costs and Cross-Country Differences in Productivity and Output.” Journal of Economic Growth 13: 145–167.

60 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles Bartel, A., C. Ichniowski, and K. Shaw. 2004. “Using ‘Insider Econometrics’ to Study Productivity.” American Economic Review 94: 217–223. Bartel, A., C. Ichniowski, and K. Shaw. 2007. “How Does Information Technology Affect Productivity? Plant-Level Comparisons of Product Innovation, Process Improvement and Worker Skills.” Quarterly Journal of Economics 122: 1721–1758. Bartel, A., C. Ichniowski, K. L. Shaw, and R. Correa. 2009. “International Differences in the Adoption and Impact of New Information Technologies and New HR Practices: The Valve- Making Industry in the United States and United Kingdom.” In International Differences in the Business Practices and Productivity of Firms, edited by R. B. Freeman and K. L. Shaw, Chapter 2, 59–78. Chicago: University of Chicago Press. Bartelmus, P. 2014. “Environmental-Economic Accounting: Progress and Digression in the SEEA Revisions.” Review of Income and Wealth 60: 887–904. Bartelmus, P. 2015. “Do We Need Ecosystem Accounts?” Ecological Economics 118: 292–298. Bartelsman, E., S. Dobbelaere, and B. Peters. 2015. “Allocation of Human Capital and Innovation at the Frontier: Firm-Level Evidence on Germany and the Netherlands.” Industrial and Corporate Change 24: 875–949. Bartelsman, E., J. Haltiwanger, and S. Scarpetta. 2009. “Measuring and Analyzing Cross-Country Differences in Firm Dynamics.” In Producer Dynamics: New Evidence from Micro Data edited by T. Dunne, J. B. Jensen, and M. J. Roberts, Chapter 1, 15–76. Chicago: University of Chicago Press. Bartelsman, E., J. Haltiwanger, and S. Scarpetta. 2013. “Cross- Country Differences in Productivity: The Role of Allocation and Selection.” American Economic Review 103: 305–334. Baumol, W. J. 1986. “Productivity Growth, Convergence, and Welfare: What the Long-Run Data Show.” American Economic Review 76: 1072–1085. Baumol, W. J., S. A. B. Blackman, and E. N. Wolff. 1989. Productivity and American Leadership: The Long View. Cambridge, MA: MIT Press. Bayo-Moriones, A., J. E. Galdon-Sanchez, and S. Martinez-de-Morentin. 2013. “The Diffusion of Pay for Performance across Occupations.” ILR Review 66: 1115–1148. Bayo-Moriones, A., and E. Huerta-Arribas. 2002. “The Adoption of Production Incentives in Spain.” British Journal of Industrial Relations 40: 709–724. Becker, G. S. 1965. “A Theory of the Allocation of Time.” Economic Journal 75: 493–517. Behrens, K., C. Ertur, and W. Koch. 2012. “‘Dual Gravity’: Using Spatial Econometrics to Control for Multilateral Resistance.” Journal of Applied Econometrics 27: 773–794. Benner, M. J., and M. Tushman. 2015. “Reflections on the 2013 Decade Award—‘Exploitation Exploration and Process Management: The Productivity Dilemma Revisited’ Ten Years Later.” Academy of Management Review 40: 497–514. Bennet, T. L. 1920. “The Theory of Measurement of Changes in Cost of Living.” Journal of the Royal Statistical Society 83: 455–462. Bergeaud, A., G. Cette, and R. Lecat. 2016. “Productivity Trends in Advanced Countries between 1890 and 2012.” Review of Income and Wealth 62: 420–444. Bergoeing, R., N. V. Loayza, and F. Piguillem. 2016. “The Whole Is Greater Than the Sum of Its Parts: Complementary Reforms to Address Microeconomic Distortions.” World Bank Economic Review 30: 268–305. Berlingieri, G., P. Blanchenay, and C. Criscuolo. 2017. “The Great Divergence(s).” OECD STI Policy Paper #39. Bernard, A. B., and C. I. Jones. 1996a. “Comparing Apples to Oranges: Productivity Convergence and Measurement across Industries and Countries.” American Economic Review 86: 1216–1238.

Overview of Productivity Analysis 61 Bernard, A. B., and C. I. Jones. 1996b. “Productivity across Industries and Countries: Time Series Theory and Evidence.” Review of Economics and Statistics 78: 135–46. Bjurek, H. 1996. “The Malmquist Total Factor Productivity Index.” Scandinavian Journal of Economics 98: 303–313. Black, S. E., and L. M. Lynch. 2004. “What’s Driving the New Economy? The Benefits of Workplace Innovation.” Economic Journal 114: F97–F116. Blancard, S., J.-P. Boussemart, W. Briec, and K. Kerstens. 2006. “Short-and Long-Run Credit Constraints in French Agriculture: A Directional Distance Function Framework Using Expenditure-Constrained Profit Functions.” American Journal of Agricultural Economics 88: 351–364. Blázquez-Gómez, L., and E. Grifell-Tatjé. 2011. “Evaluating the Regulator: Winners and Losers in the Regulation of Spanish Electricity Distribution.” Energy Economics 33: 807–815. Bloom, N., E. Brynjolfsson, L. Foster, R. Jarmin, M. Patnaik, I. Saporta-Eksten, and J. Van Reenen. 2014. “IT and Management in America.” CEP Discussion Paper No 1258, Centre for Economic Performance, London School of Economics and Political Science. http://cep.lse. ac.uk/pubs/download/dp1258.pdf. Bloom, N., C. Genakos, R. Sadun, and J. Van Reenen. 2012. “Management Practices Across Firms and Countries.” Academy of Management Perspectives 26: 12–33. Bloom, N., R. Lemos, R. Sadun, D. Scur, and J. Van Reenen. 2014. “The New Empirical Economics of Management.” Journal of the European Economic Association 12: 835–876. Bloom, N., R. Sadun, and J. Van Reenen. 2016. “Management as a Technology?” National Bureau of Economic Research. http://www.nber.org/papers/w22327.pdf. Bloom, N., and J. Van Reenen. 2007. “Measuring and Explaining Management Practices Across Firms and Countries.” Quarterly Journal of Economics 122: 1351–1408. Bos, J. W. B., R. C. R. van Lamoen, and M. W. J. L. Sanders. 2016. “Producing Innovations: Determinants of Innovativity and Efficiency.” In Advances in Efficiency and Productivity, edited by J. Aparicio, C. A. K. Lovell, and J. T. Pastor, Chapter 10, 227–248. New York: Springer. Boskin, M. J., E. Dulberger, R. Gordon, Z. Griliches, and D. Jorgenson. 1996. “Towards a More Accurate Measure of the Cost of Living.” Final Report to the US Senate Finance Committee. https://www.ssa.gov/history/reports/boskinrpt.html. Bottazzi, G., A. Secchi, and F. Tamagni. 2008. “Productivity, Profitability and Financial Performance.” Industrial and Corporate Change 17: 711–751. Bourlès, R., G. Cette, J. Lopez, J. Mairesse, and G. Nicoletti. 2013. “Do Product Market Regulations in Upstream Sectors Curb Productivity Growth? Panel Data Evidence for OECD Countries.” Review of Economics and Statistics 95: 1750–1768. Bournakis, I., M. Vecchi, and F. Venturini. 2018. “Off-Shoring, Specialization and R&D.” Review of Income and Wealth, 64: 26–51. doi:10.1111.roiw.12239. Brandenburger, A. M., and H. W. Stuart, Jr. 1996. “Value-Based Business Strategy.” Journal of Economics and Management Strategy 5: 5–24. Brandt, N., P. Schreyer, and V. Zipperer. 2017. “Productivity Measurement with Natural Capital.” Review of Income and Wealth 63: S7–S21. Brea- Solís, H., R. Casadesus- Masanell, and E. Grifell- Tatjé. 2015. “Business Model Evaluation: Quantifying Walmart’s Sources of Advantage.” Strategic Entrepreneurship Journal 9: 12–33. Bridgman, B. 2016. “Accounting for Household Production in the National Accounts: An Update, 1965–2014.” Survey of Current Business 96: 1–5.

62 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles Bryan, L. L. 2007. “The New Metrics of Corporate Performance: Profit per Employee.” McKinsey Quarterly. http://www.mckinsey.com/business-functions/strategy-and-corporate-finance/ our-insights/the-new-metrics-of-corporate-performance-profit-per-employee. Brynjolfsson, E., and A. McAfee. 2014. The Second Machine Age: Work, Progress and Prosperity in a Time of Brilliant Technologies. New York: W. W. Norton. Burkhart, M., F. Panunzi, and A. Shleifer. 2003. “Family Firms.” Journal of Finance 58: 2167–2201. Bushnell, J. B., and C. Wolfram. 2009. “The Guy at the Controls: Labor Quality and Power Plant Efficiency.” In International Differences in the Business Practices and Productivity of Firms, edited by R. B. Freeman and K. L. Shaw, Chapter 3, 79–102. Chicago: University of Chicago Press. Caballero, R., T. Hoshi, and A. K. Kashyap. 2008. “Zombie Lending and Depressed Restructuring in Japan.” American Economic Review 98: 1943–1977. Cameron, G., J. Proudman, and S. Redding. 1999. “Productivity Convergence and International Openness.” In Openness and Growth, edited by J. Proudman and S. Redding, 221–260. London: Bank of England. Carrington, R., N. Puthucheary, D. Rose, and S. Yaisawarng. 1997. “Performance Measurement in Government Service Provision: The Case of Police Services in New South Wales.” Journal of Productivity Analysis 8: 415–430. Caselli, F., and N. Gennaioli. 2013. “Dynastic Management.” Economic Inquiry 51: 971–996. Caves, D. W., L. R. Christensen, and W. E. Diewert. 1982. “The Economic Theory of Index Numbers and the Measurement of Input, Output, and Productivity.” Econometrica 50: 1393–1414. Caves, D. W., L. R. Christensen, and J. A. Swanson. 1980. “Productivity in U.S. Railroads, 1951– 1974.” Bell Journal of Economics 11: 166–181. CERC (Centre d’Études des Revenus et des Coûts). 1969. “ ‘Surplus de productivité globale’ et ‘Comptes de surplus.’ ” Documents du Centre d’Études des Revenus et des Coûts, No. 1, 1er trimestre. Paris: CERC. Cette, G., J. Lopez, and J. Mairesse. 2016. “Market Regulations, Prices and Productivity.” American Economic Review 106: 104–108. Cette, G., J. Lopez, and J. Mairesse. 2017. “Upstream Product Market Regulations, ICT, R&D and Productivity.” Review of Income and Wealth 63: S68–S89. Chandler, A. D. 1962. Strategy and Structure: Chapters on the History of the Industrial Enterprise. Cambridge, MA: MIT Press. Chang, C.- C., and S. E. Stefanou. 1988. “Specification and Estimation of Asymmetric Adjustment Rates for Quasi-Fixed Factors of Production.” Journal of Economic Dynamics and Control 12: 145–51. Charnes, A., W. W. Cooper, and E. Rhodes. 1978. “Measuring the Efficiency of Decision- Making Units.” European Journal of Operational Research 2: 429–444. Cherchye, L., W. Moesen, N. Rogge, and T. Van Puyenbroeck. 2007. “An Introduction to ‘Benefit of the Doubt’ Composite Indicators.” Social Indicators Research 82: 111–145. Clark, C. 1940. The Conditions of Economic Progress. London: Macmillan. Coad, A. 2007. “Testing the Principle of ‘Growth of the Fitter’: The Relationship Between Profits and Firm Growth.” Structural Change and Economic Dynamics 18: 370–386. Collard-Wexler, A., and J. De Loecker. 2015. “Reallocation and Technology: Evidence from the US Steel Industry.” American Economic Review 105: 131–171. Corrado, C., J. Haltiwanger, and D. Sichel. 2005. “Introduction.” In Measuring Capital in the New Economy, edited by C. Corrado, J. Haltiwanger, and D. Sichel, 1–10. Chicago: University of Chicago Press.

Overview of Productivity Analysis 63 Corrado, C., and C. Hulten. 2014. “Innovation Accounting.” In Measuring Economic Sustainability and Progress, edited by D. W. Jorgenson, S. Landefeld, and P. Schreyer, Chapter 18, 595–628. Chicago: University of Chicago Press. Corrado, C., C. Hulten, and D. Sichel. 2005. “Measuring Capital and Technology: An Expanded Framework.” In Measuring Capital in the New Economy, edited by C. Corrado, J. Haltiwanger, and D. Sichel, Chapter 1, 11–46. Chicago: University of Chicago Press. Corrado, C., C. Hulten, and D. Sichel. 2009. “Intangible Capital and U.S. Economic Growth.” Review of Income and Wealth 55: 661–685. Crafts, N. 2004. “Steam as a General Purpose Technology: A Growth Accounting Perspective.” Economic Journal 114: 338–351. Crafts, N., and K. O’Rourke. 2013. “Twentieth Century Growth.” Oxford University Economic and Social History Series 117. http://www.economics.ox.ac.uk/materials/papers/12884/ Crafts%20O’Rourke%20117.pdf. Dakpo, K. H., P. Jeanneaux, and L. Latruffe. 2016. “Modelling Pollution-Generating Technologies in Performance Benchmarking: Recent Developments, Limits and Future Prospects in the Nonparametric Framework.” European Journal of Operational Research 250: 347–359. David, P. A. 1990. “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox.” American Economic Review 80: 355–361. Davis, H. S. 1947. The Industrial Study of Economic Progress. Philadelphia: University of Pennsylvania Press. Davis, H. S. 1955. Productivity Accounting. Philadelphia: University of Pennsylvania Press. De Backer, K., C. Menon, I. Desnoyers-James, and L. Moussiegt. 2016. “Reshoring: Myth or Reality?” OECD Science, Technology and Industry Policy Papers No. 27. http://dx.doi.org/ 10.1787/5jm56frbm38s-en. Debreu, G. 1951. “The Coefficient of Resource Utilization.” Econometrica 19: 273–292. Dedrick, J., K. L. Kraemer, and G. Linden. 2010. “Who Profits from Innovation in Global Value Chains? A Study of the iPod and Notebook PCs.” Industrial and Corporate Change 19: 81–116. De Loecker, J., P. K. Goldberg, A. K. Khandelwal, and N. Pavcnik. 2016. “Prices, Markups and Trade Reform.” Econometrica 84: 445–510. De Loecker, J., and F. Warzynski. 2012. “Markups and Firm-Level Export Status.” American Economic Review 102: 2437–2471. Denison, E. F. 1962. The Sources of Economic Growth in the United States and the Alternatives Before Us. Supplementary Paper No. 13, Committee for Economic Development. Denison, E. F. 1974. Accounting for United States Economic Growth 1929–1969. Washington, DC: The Brookings Institution. D’Erasmo, P. N., and H. J. Moscoso Boedo. 2012. “Financial Structure, Informality and Development.” Journal of Monetary Economics 59: 286–302. Diewert, W. E. 1992. “Fisher Ideal Output, Input and Productivity Indexes Revisited.” Journal of Productivity Analysis 3: 211–248. Diewert, W. E. 2014. “US TFP Growth and the Contribution of Changes in Export and Import Prices to Real Income Growth.” Journal of Productivity Analysis 41: 19–39. Diewert, W. E., and C. J. Morrison. 1986. “Adjusting Output and Productivity Indexes for Changes in the Terms of Trade.” Economic Journal 96: 659–679. Domar, E. D. 1966. “The Soviet Collective Farm as a Producer Cooperative.” American Economic Review 56: 734–757. Dowrick, S., and D.-T. Nguyen. 1989. “OECD Comparative Economic Growth 1950–85: Catch- up and Convergence.” American Economic Review 79: 1010–1030.

64 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles Draca, M., R. Sadun, and J. Van Reenen. 2007. “Productivity and ICTs: A Review of the Evidence.” In The Oxford Handbook of Information and Communication Technologies, edited by R. Mansell, C. Avgerou, D. Quah, and R. Silverstone, Chapter 5, 100–147. Oxford: Oxford University Press. Drucker, P. F. 1954. The Practice of Management. New York: Harper & Row. Druska, V., and W. C. Horrace. 2004. “Generalized Moments Estimation for Spatial Panel Data: Indonesian Rice Farming.” American Journal of Agricultural Economics 86: 185–198. Easterly, W., and R. Levine. 2001. “It’s Not Factor Accumulation: Stylized Facts and Growth Models.” World Bank Economic Review 15: 177–219. Economist. 2016a. “The Problem with Profits” and “Too Much of a Good Thing.” The Economist, March 26. http://www.economist.com/printedition/2016-03-26. Economist. 2016b. “Special Report: Artificial Intelligence.” The Economist, June 25. http://www. economist.com/printedition/2016-06-25. Economist. 2016c. “A Sea of Black Mirrors: The Niche in Phones That Are Different.” The Economist, November 5. http://www.economist.com/printedition/2016-11-05. Economist. 2017. “Every Little Helps.” The Economist, January 14. http://economist.com/ printedition/2017-01-14. Égert, B. 2016. “Regulation, Institutions and Productivity: New Macroeconomic Evidence from OECD Countries.” American Economic Review 106: 109–113. Égert, B., and I. Wanner. 2016. “Regulations in Services Sectors and Their Impact on Downstream Industries: The OECD 2013 Regimpact Indicator.” OECD Economics Department Working Papers No. 1303. http://dx.doi.org/10.1787/5jlwz7kz39q8-en. Eldor, D., and E. Sudit. 1981. “Productivity-Based Financial Net Income Analysis.” Omega 9: 605–611. Ertur, C., and W. W. Koch. 2007. “Growth, Technological Interdependence and Spatial Externalities: Theory and Evidence.” Journal of Applied Econometrics 22: 1033–1062. Eslava, M., J. Haltiwanger, A. Kugler, and M. Kugler. 2004. “The Effects of Structural Reforms on Productivity and Profitability Enhancing Reallocation: Evidence from Colombia.” Journal of Development Economics 75: 333–371. Eslava, M., J. Haltiwanger, A. Kugler, and M. Kugler. 2010. “Factor Adjustments after Deregulation: Panel Evidence From Colombian Plants.” Review of Economics and Statistics 92: 378–391. Eslava, M., J. Haltiwanger, A. Kugler, and M. Kugler. 2013. “Trade and Market Selection: Evidence from Manufacturing Plants in Colombia.” Review of Economic Dynamics 16: 135–158. Fabricant, S. 1940. The Output of Manufacturing Industries, 1899–1937. New York: National Bureau of Economic Research. http://www.nber.org/chapters/c6435. Fabricant, S. 1961. “Basic Facts on Productivity Change.” In Productivity Trends in the United States, edited by J. W. Kendrick, xxxv–lii. Princeton, NJ: Princeton University Press. Also issued as National Bureau of Economic Research Occasional Paper 63, http://papers.nber. org/books/fabr59-1. Färe, R., and S. Grosskopf. 1994. Cost and Revenue Constrained Production. New York: Springer-Verlag. Färe, R., and S. Grosskopf .1996. Intertemporal Production Frontiers: With Dynamic DEA. Boston: Kluwer Academic. Felipe, J., and J. McCombie. 2017. “The Debate about the Sources of Growth in East Asia after a Quarter of a Century: Much Ado about Nothing.” Asian Development Bank Economics

Overview of Productivity Analysis 65 Working Paper Series No. 512, Manila, Philippines. https://www.adb.org/publications/ debate-sources-growth-east-asia. Fernandes, A. P., and P. Ferreira. 2017. “Financing Constraints and Fixed- Term Employment: Evidence from the 2008-9 Financial Crisis.” European Economic Review 92: 215–238. Ferrando, A., and A. Ruggieri. 2015. “Financial Constraints and Productivity: Evidence from Euro Area Companies.” European Central Bank Working Paper No. 1823. http://www.ecb. europa.eu/pub/pdf/scpwps/ecbwp1823.en.pdf. Fisher, S. R., and M. A. White. 2000. “Downsizing in a Learning Organization: Are There Hidden Costs?” Academy of Management Review 25: 244–251. Fixler, D., and D. S. Johnson. 2014. “Accounting for the Distribution of Income in the US National Accounts.” In Measuring Economic Sustainability and Progress, edited by D. W. Jorgenson, J. S. Landefeld, and P. Schreyer, Chapter 8, 213–244. Chicago: University of Chicago Press. Førsund, F. R. 2009. “Good Modelling of Bad Outputs: Pollution and Multiple-Output Production.” International Review of Environmental and Resource Economics 3: 1–38. Fort, T. C. 2017. “Technology and Production Fragmentation: Domestic versus Foreign Sourcing.” Review of Economic Studies 84: 650–687. Foster, L., J. Haltiwanger, and C. J. Krizan. 2001. “Aggregate Productivity Growth: Lessons from Microeconomic Evidence.” In New Developments in Productivity Analysis, edited by C. R. Hulten, E. R. Dean, and M. J. Harper, Chapter 8, 303–372. Chicago: University of Chicago Press. Foster, L., J. Haltiwanger, and C. J. Krizan. 2006. “Market Selection, Reallocation, and Restructuring in the U.S. Retail Trade Sector in the 1990s.” Review of Economics and Statistics 88: 748–758. Foster, L., J. Haltiwanger, and C. Syverson. 2008. “Reallocation, Firm Turnover, and Efficiency: Selection on Productivity or Profitability?” American Economic Review 98: 394–425. Freeman, R., and J. Medoff. 1984. What Do Unions Do? New York: Basic Books. Frey, C. B., and M. A. Osborne. 2013. “The Future of Employment: How Susceptible Are Jobs to Computerisation?” http://www.oxfordmartin.ox.ac.uk/downloads/academic/The_Future_ of_Employment.pdf. Frisch, R. 1965. Theory of Production. Dordrecht: D. Reidel. Gandolfi, F., and M. Hansson. 2011. “Causes and Consequences of Downsizing: Towards an Integrative Framework.” Journal of Management & Organization 17: 498–521. Garvy, G. 1954. “Functional and Size Distributions of Income and Their Meaning.” American Economic Review 44: 236–253. Gennaioli, N., R. La Porta, F. Lopez-de-Silanes, and A. Shleifer. 2013. “Human Capital and Regional Development.” Quarterly Journal of Economics 128: 105–164. Gennaioli, N., R. La Porta, F. Lopez-de-Silanes, and A. Shleifer. 2014. “Growth in Regions.” Journal of Economic Growth 19: 259–309. Gibbons, R., and R. Henderson. 2012a. “What Do Managers Do? Exploring Persistent Performance Differences among Seemingly Similar Enterprises.” In The Handbook of Organizational Economics, edited by R. Gibbons and J. Roberts, Chapter 17, 680–731. Princeton, NJ: Princeton University Press. Gibbons, R., and R. Henderson. 2012b. “Relational Contracts and Organizational Capabilities.” Organization Science 23: 1350–1364.

66 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles Glass, A., K. Kenjegalieva, and J. Paez-Farrell. 2013. “Productivity Growth Decomposition Using a Spatial Autoregressive Frontier Model.” Economics Letters 119: 291–295. Glass, A. J., K. Kenjegalieva, and R. C. Sickles. 2016a. “Returns to Scale and Curvature in the Presence of Spillovers: Evidence from European Countries.” Oxford Economic Papers 68: 40–63. Glass, A. J., K. Kenjegalieva, and R. C. Sickles. 2016b. “A Spatial Autoregressive Stochastic Frontier Model for Panel Data with Asymmetric Efficiency Spillovers.” Journal of Econometrics 190: 289–300. Goesaert, T., M. Heinz, and S. Vanormelingen. 2015. “Downsizing and Firm Performance: Evidence from German Firm Data.” Industrial and Corporate Change 24: 1443–1472. Goldin, C., and L. F. Katz. 1998. “The Origins of Technology-Skill Complementarity.” Quarterly Journal of Economics 113: 693–732. Gordon, R. J. 2016. The Rise and Fall of American Growth: The U.S. Standard of Living since the Civil War. Princeton, NJ: Princeton University Press. Grifell- Tatjé, E., and C. A. K. Lovell. 2004. “Decomposing the Dividend.” Journal of Comparative Economics 32: 500–518. Grifell-Tatjé, E., and C. A. K. Lovell. 2008. “Productivity at the Post: Its Drivers and Its Distribution.” Journal of Regulatory Economics 33: 133–158. Grifell-Tatjé, E., and C. A. K. Lovell. 2013. “Advances in Cost Frontier Analysis of the Firm.” In The Oxford Handbook of Managerial Economics, edited by C. R. Thomas and W. F. Shughart II, Chapter 4, 66–88. New York: Oxford University Press. Grifell-Tatjé, E., and C. A. K. Lovell. 2015. Productivity Accounting: The Economics of Business Performance. New York: Cambridge University Press. Grifell-Tatjé, E., C. A. K. Lovell, and P. Turon. 2016. “Social Economic Progress and the Analysis of Business Behaviour.” Working Paper. Griliches, Z. 1957. “Hybrid Corn: An Exploration in the Economics of Technological Change.” Econometrica 25: 501–522. Griliches, Z. 1969. “Capital-Skill Complementarity.” Review of Economics & Statistics 51: 465–468. Griliches, Z. 1992. Output Measurement in the Service Sectors. Chicago: University of Chicago Press. Griliches, Z. 1994. “Productivity, R&D, and the Data Constraint.” American Economic Review 84: 1–23. Grosskopf, S., K. J. Hayes, L. L. Taylor, and W. L. Weber. 1997. “Budget-Constrained Frontier Measures of Fiscal Equality and Efficiency in Schooling.” Review of Economics and Statistics 79: 116–124. Guthrie, J. P., and D. K. Datta. 2008. “Dumb and Dumber: The Impact of Downsizing on Firm Performance as Moderated by Industry Conditions.” Organization Science 19: 108–123. Hall, B. 2011. “Innovation and Productivity.” Nordic Economic Policy Review 2: 167–204. Hall, R. E., and C. I. Jones. 1999. “Why Do Some Countries Produce So Much More Output Per Worker Than Others?” Quarterly Journal of Economics 114: 83–116. Haltiwanger, J. 2016. “Firm Dynamics and Productivity: TFPQ, TFPR and Demand Side Factors.” Economía 17. http://www.cid.harvard.edu/Economia/contents.htm. Hamermesh, D. S., and G. A. Pfann. 1996. “Adjustment Costs in Factor Demand.” Journal of Economic Literature 34: 1264–1292. Han, J., D. Ryu, and R. C. Sickles. 2016. “How to Measure Spillover Effects of Public Capital Stock: A Spatial Autoregressive Stochastic Frontier Model.” Advances in Econometrics 37: 259–294.

Overview of Productivity Analysis 67 Helper, S., and R. Henderso.n 2014. “Management Practices, Relational Contracts, and the Decline of General Motors.” Journal of Economic Perspectives 28: 49–72. Hicks, J. R. 1961. “Measurement of Capital in Relation to the Measurement of Economic Aggregates.” In The Theory of Capital, edited by F. A. Lutz and D. C. Hague, 18–31. London: Macmillan. Hoch, I. 1955. “Estimation of Production Function Parameters and Testing for Efficiency.” Econometrica 23: 325–326. Hopenhayn, H. A. 2014. “Firms, Misallocation and Aggregate Productivity: A Review.” Annual Review of Economics 6: 735–770. Hsieh, C.-T. 2002. “What Explains the Industrial Revolution in East Asia? Evidence from the Factor Markets.” American Economic Review 92: 502–526. Hsieh, C.-T. 2015. “Policies for Productivity Growth.” OECD Productivity Working Papers No. 03. http://www.oecd-ilibrary.org/economics/policies-for-productivity-growth_5jrp1f5rddtc-en. Hsieh, C.-T., and P. J. Klenow. 2009. “Misallocation and Manufacturing TFP in China and India.” Quarterly Journal of Economics 124: 1403–1448. Hsieh, C.-T., and E. Moretti. 2015. “Why Do Cities Matter? Local Growth and Aggregate Growth.” National Bureau of Economic Research Working Paper 21154. http://www.nber. org/papers/w21154. Hultberg, P. T., M. I. Nadiri, and R. C. Sickles. 2004. “Cross-Country Catch-up in the Manufacturing Sector: Impacts of Heterogeneity on Convergence and Technology Adoption.” Empirical Economics 29: 753–68. Hultberg, P. T., M. I. Nadiri, R. C. Sickles and P. T. Hultberg. 1999. “An International Comparison of Technology Adoption and Efficiency: A Dynamic Panel Model.” Annales d’Économie et de Statistique 55–56: 449–474. Hulten, C. R. 2001. “Total Factor Productivity: A Short Biography.” In New Developments in Productivity Analysis, edited by C. R. Hulten, E. R. Dean, and M. J. Harper, Chapter 1, 1–53. Chicago: University of Chicago Press. Iacovone, L., and G. A. Crespi. 2010. “Catching up with the Technological Frontier: Micro- Level Evidence on Growth and Convergence.” Industrial and Corporate Change 19: 2073–2096. Ichniowski, C., and K. Shaw. 2012. “Insider Econometrics.” In The Handbook of Organizational Economics, edited by R. Gibbons and J. Roberts, Chapter 7, 263–311. Princeton, NJ: Princeton University Press. Ichniowski, C., K. Shaw, and G. Prennushi. 1997. “The Effects of Human Resource Management Practices on Productivity: A Study of Steel Finishing Lines.” American Economic Review 87: 291–313. Inklaar, R., and W. E. Diewert. 2016. “Measuring Industry Productivity and Cross-Country Convergence.” Journal of Econometrics 191: 426–433. Inklaar, R., and M. P. Timmer. 2009. “Productivity Convergence across Industries and Countries: The Importance of Theory-Based Measurement.” Macroeconomic Dynamics 13 (Suppl 2): 218–240. Johnson, H. T. 1975. “Management Accounting in an Early Integrated Industrial: E. I. duPont de Nemours Powder Company, 1903–1912.” Business History Review 49: 184–204. Johnson, H. T. 1978. “Management Accounting in an Early Multidivisional Organization: General Motors in the 1920s.” Business History Review 52: 490–517. Jorgenson, D. W., and Z. Griliches. 1967. “The Explanation of Productivity Change.” Review of Economic Studies 34: 249–283.

68 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles Kendrick, J. W. 1961. Productivity Trends in the United States. Princeton, NJ: Princeton University Press. Kendrick, J. W. 1984. Improving Company Productivity: Handbook with Case Studies. Baltimore, MD: Johns Hopkins University Press. Kendrick, J. W., and D. Creamer. 1961. Measuring Company Productivity: Handbook with Case Studies. Studies in Business Economics 74. New York: The Conference Board. Kim, J.-I., and L. J. Lau. 1994. “The Sources of Economic Growth of the East Asian Newly Industrialized Countries.” Journal of the Japanese and International Economies 8: 235–7 1. Krueger, A. B. 2013. “Land of Hope and Dreams: Rock and Roll, Economics and Rebuilding the Middle Class.” www.whitehouse.gov/sites/default/files/docs/hope_and_dreams_-_final.pdf. Krugman, P. 1994. “The Myth of Asia’s Miracle.” Foreign Affairs 73: 62–78. Krussell, P., L. E. Ohanian, J.- V. Ríos- Rull, and G. L. Violante. 2000. “Capital- Skill Complementarity and Inequality: A Macroeconomic Analysis.” Econometrica 68: 1029–1053. Kumar, S., and R. R. Russell. 2002. “Technological Change, Technological Catch-up, and Capital Deepening: Relative Contributions to Growth and Convergence,” American Economic Review 92: 527–548. Kumbhakar, S. C., and E. Malikov. 2018. “Good Modeling of Bad Outputs: Editors’ Introduction.” Empirical Economics 54: 1–6. http://dx.doi.rg/10.1007/s00181-017-1231-8. Lakner, C., and B. Milanovic. 2016. “Global Income Distribution: From the Fall of the Berlin Wall to the Great Recession.” World Bank Economic Review 30: 203–232. Landes, D. S. (.1990. “Why Are We So Rich and They So Poor?” American Economic Review 80: 1–13. Lansink, A. O., S. E. Stefanou, and T. Serra. 2015. “Primal and Dual Dynamic Luenberger Productivity Indicators.” European Journal of Operational Research 241, 555–563. La Porta, R., and A. Shleifer. 2014. “Informality and Development.” Journal of Economic Perspectives 28: 109–126. Lauwers, L. 2009. “Justifying the Incorporation of the Materials Balance Principle into Frontier-Based Eco-Efficiency Models.” Ecological Economics 68: 1605–1614. Lawrence, D., W. E. Diewert, and K. J. Fox. 2006. “The Contributions of Productivity, Price Change and Firm Size to Profitability.” Journal of Productivity Analysis 26: 1–13. Lazear, E. 2000. “Performance Pay and Productivity.” American Economic Review 90: 1346–1361. Lewis, W. W. 2004. The Power of Productivity. Chicago: University of Chicago Press. Lieberman, M. B., R. Garcia-Castro, and N. Balasubramanian. 2017. “Measuring Value Creation and Appropriation in Firms: The VCA Model.” Strategic Management Journal 38: 1193–1211. Lucas, R. E., Jr. 1988. “On the Mechanics of Economic Development.” Journal of Monetary Economics 22: 3–42. Luh, Y.-H., and S. E. Stefanou. 1991. “Productivity Growth in US Agriculture under Dynamic Adjustment.” American Journal of Agricultural Economics 73: 1116–25. Mairesse, J., and P. Mohnen. 2002. “Accounting for Innovation and Measuring Innovativeness: An Illustrative Framework and an Application.” American Economic Review 92: 226–230. Malmquist, S. 1953. “Index Numbers and Indifference Surfaces.” Trabajos de Estadistica 4: 209–242. Mankiw, N. G., D. Rome, and D. N. Weil. 1992. “A Contribution to the Empirics of Economic Growth.” Quarterly Journal of Economics 107: 407–437. Massell, B. F. 1967. “Elimination of Management Bias from Production Functions Fitted to Cross-Section Data: A Model and an Application to African Agriculture.” Econometrica 35: 495–508.

Overview of Productivity Analysis 69 McKinsey Global Institute. 2016. “Poorer Than Their Parents? A New Perspective on Income Inequality.” http://www.mckinsey.com/global-themes/employment-and-growth/ poorer-than-their-parents-a-new-perspective-on-income-inequality. Meeusen, W., and J. van den Broeck. 1977. “Efficiency Estimation from Cobb-Douglas Production Functions with Composed Error.” International Economic Review 18: 435–444. Midrigan, V., and D. Y. Xu. 2014. “Finance and Misallocation: Evidence from Plant-Level Data.” American Economic Review 104: 422–458. Milgrom, P., and J. Roberts. 1990. “The Economics of Modern Manufacturing.” American Economic Review 80: 511–528. Milgrom, P., and J. Roberts. 1995. “Complementarities and Fit: Strategy, Structure and Organizational Change in Manufacturing.” Journal of Accounting and Economics 19: 179–208. Miller, D. M. 1984. “Profitability = Productivity + Price Recovery.” Harvard Business Review 62: 145–153. Mills, F. C. 1932. Economic Tendencies in the United States: Aspects of Pre-War and Post-War Changes. Cambridge, MA: National Bureau of Economic Research. http://papers.nber.org/ books/mill32-1. Mohnen, P., and B. H. Hall. 2013. “Innovation and Productivity: An Update.” Eurasian Business Review 3: 47–65. Mokyr, J., C. Vickers, and N. L. Ziebarth. 2015. “The History of Technological Anxiety and the Future of Economic Growth: Is This Time Different?” Journal of Economic Perspectives 29: 31–50. Moll, B. 2014. “Productivity Losses from Financial Frictions: Can Self-Financing Undo Capital Misallocation?” American Economic Review 104: 3186–3221. Moorsteen, R. H. 1961. “On Measuring Productive Potential and Relative Efficiency.” Quarterly Journal of Economics 75: 451–467. Moscoso Boedo, H. J., and T. Mukoyama. 2012. “Evaluating the Effects of Entry Regulations and Firing Costs on International Income Differences.” Journal of Economic Growth 17: 143–170. Mundlak, Y. 1961. “Empirical Production Function Free of Management Bias.” Journal of Farm Economics 43: 44–56. Nelson, R. R., and S. G. Winter. 1982. An Evolutionary Theory of Economic Change. Cambridge, MA: Harvard University Press. Nelson, R. R., and G. Wright. 1992. “The Rise and Fall of American Technological Leadership: The Postwar Era in Historical Perspective.” Journal of Economic Literature 30: 1931–64. Niebel, T., M. O’Mahoney, and M. Saam. 2017. “The Contribution of Intangible Assets to Sectoral Productivity Growth in the EU.” Review of Income and Wealth 63: S49–S67. Nordhaus, W. D. 2006. “Geography and Macroeconomics: New Data and New Findings.” Proceedings of the National Academy of Sciences 103: 3510–3517. Nordhaus, W. D., and E. C. Kokkelenberg. 1999. Nature’s Numbers: Expanding the National Accounts to Include the Environment. Washington, DC: National Academy Press. Nordhaus, W. D., and J. Tobin. 1972. “Is Growth Obsolete?” In Economic Research: Retrospect and Prospect, Vol. 5: Economic Growth edited by William D. Nordhaus and James Tobin, Chapter 1, 1–80. Cambridge, MA: National Bureau of Economic Research http://www.nber. org/chapters/c7620. North, D. C. 1990. Institutions, Institutional Change and Economic Performance. Cambridge: Cambridge University Press. O’Donnell, C. J. 2012. “An Aggregate Quantity Framework for Measuring and Decomposing Productivity Change.” Journal of Productivity Analysis 38: 255–272.

70 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles OECD. 2001a. Measuring Productivity: OECD Manual. http://www.oecd.org/std/productivity- stats/2352458.pdf OECD. 2001b. “Productivity and Firm Dynamics: Evidence from Microdata.” In OECD Economic Outlook 69: Chapter 7, 209–223. www.oecd.org/eco/outlook/2079019.pdf. OECD. 2002. Measuring the Non-Observed Economy: A Handbook. https://www.oecd.org/std/ na/1963116.pdf. OECD. 2011. “Incorporating Estimates of Household Production of Non-Market Services into International Comparisons of Material Well- Being.” www.oecd.org/officialdocuments/ publicdisplaydocumentpdf/?cote=std/doc(2011)7&doclanguage=en. OECD. 2014. “All on Board. Making Inclusive Growth Happen.” http://www.oecd.org/ inclusive-growth/All-on-Board-Making-Inclusive-Growth-Happen.pdf. OECD. 2015a. Economic Policy Reforms: Going For Growth. http://www.oecd-ilibrary.org/economics/economic-policy-reforms-2015_growth-2015-en. OECD. 2015b. The Future of Productivity. http://www.oecd-ilibrary.org/economics/the-future- of-productivity_9789264248533-en. OECD. 2016a. The Productivity– Inclusiveness Nexus. Meeting of the OECD Council at Ministerial Level, Paris, June 1–2. https://www.oecd.org/global-forum-productivity/library/ The-Productivity-Inclusiveness-Nexus-Preliminary.pdf. OECD. 2016b. “Promoting Productivity and Equality: A Twin Challenge.” In OECD Economic Outlook 2016(1), Chapter 2. http://www.oecd.org/eco/outlook/OECD-Economic-Outlook- June-2016-promoting-productivity-and-equality.pdf. Oh, Y., D.-h. Oh, and J. D. Lee. Forthcoming. “A Sequential Global Malmquist Productivity Index: Productivity Growth Index for Unbalanced Panel Data Considering the Progressive Nature of Technology.” Empirical Economics 52: 1651–1674. http://dx.doi.rg/10.1007/s00181-016-1104-6. Olley, S. G., and A. Pakes. 1996. “The Dynamics of Productivity in the Telecommunications Equipment Industry.” Econometrica 64: 1263–1297. Olsen, K. B. 2006. “Productivity Impacts of Offshoring and Outsourcing: A Review,” OECD Science, Technology and Industry Working Papers, 2006/01. http://dx.doi.org/10.1787/ 685237388034. Parente, S. L., and E. C. Prescott. 1993. “Changes in the Wealth of Nations.” Federal Reserve Bank of Minneapolis Quarterly Review 17: 3–16. Pérez-López, G., D. Prior, J. L. Zafra-Gómez, and A. M. Plata-Díaz. 2016. “Cost Efficiency in Municipal Solid Waste Service Delivery. Alternative Management Forms in Relation to Local Population Size.” European Journal of Operational Research 255: 583–592. Peyrache, A., and A. Zago. 2016. “Large Courts, Small Justice! The Inefficiency and the Optimal Structure of the Italian Justice Sector.” Omega 64: 42–56. Poissonnier, A., and D. Roy. 2017. “Household Satellite Account for France: Methodological Issues on the Assessment of Domestic Production.” Review of Income and Wealth, 63: 353–377. doi: 10.1111/roiw.12216. Porter, M. E. 1991. “America’s Green Strategy.” Scientific American (April): 168. Raymond, W., J. Mairesse, P. Mohnen, and F. Palm. 2015. “Dynamic Models of R&D, Innovation and Productivity: Panel Data Evidence from Dutch and French Manufacturing.” European Economic Review 78: 285–306. Reijnders, L. S. M., M. P. Timmer, and X. Ye. 2016. “Offshoring, Biased Technical Change and Labor Demand: New Evidence from Global Value Chains.” GGDC Research Memorandum 164, Groningen Growth and Development Centre, University of Groningen. http://www. ggdc.net/publications/memorandum/gd164.pdf.

Overview of Productivity Analysis 71 Rodrick, D. 2013. “Unconditional Convergence in Manufacturing.” Quarterly Journal of Economics 128: 165–204. Romer, P. M. 1986. “Increasing Returns and Long-Run Growth.” Journal of Political Economy 94: 1002–1037. Royal Swedish Academy of Sciences. 1992. “Press Release.” https://www.nobelprize.org/nobel_ prizes/economic-sciences/laureates/1992/press.html Royal Swedish Academy of Sciences. 1993. “Press Release.” https://www.nobelprize.org/nobel_ prizes/economic-sciences/laureates/1993/press.html Salter, W. E. G. 1960. Productivity and Technical Change. Cambridge: Cambridge University Press. Schenkel, A., and R. Teigland. 2017. “Why Doesn’t Downsizing Deliver? A Multi-Level Model Integrating Downsizing, Social Capital, Dynamic Capabilities and Firm Performance.” International Journal of Human Resource Management 28: 1065–1107. Schmookler, J. 1952. “The Changing Efficiency of the American Economy, 1869–1938.” Review of Economics and Statistics 34: 214–231. Schreyer, P. 2012. “Output, Outcome and Quality Adjustment in Measuring Health and Education Services.” Review of Income and Wealth 58: 257–278. Schreyer, P., and W. E. Diewert. 2014. “Household Production, Leisure and Living Standards.” In Measuring Economic Sustainability and Progress, edited by D. W. Jorgenson, J. S. Landefeld, and P. Schreyer, Chapter 4, 89–114. Chicago: University of Chicago Press. Schultz, T. W. 1964. Transforming Traditional Agriculture. New Haven, CT: Yale University Press. Schultz, T. W. 1975. “The Value of the Ability to Deal with Disequilibria.” Journal of Economic Literature 13: 827–846. Schumpeter, J. A. 1942. Capitalism, Socialism and Democracy. New York: Harper & Row. Shaw, K. 2009. “Insider Econometrics: A Roadmap with Stops along the Way.” Labour Economics 16: 607–617. Shephard, R. W. 1953. Cost and Production Functions. Princeton, NJ: Princeton University Press. Shephard, R. W. 1970. Theory of Cost and Production Functions. Princeton, NJ: Princeton University Press. Shephard, R. W. 1974. Indirect Production Functions. Meisenheim Am Glan: Verlag Anton Hain. Shephard, R. W., and R. Färe. 1980. Dynamic Theory of Production Correspondences. Cambridge, MA: Oelgeschlager, Gunn and Hain Publishers. Shestalova, V. 2003. “Sequential Malmquist Indices of Productivity Growth: An Application to OECD Industrial Activities.” Journal of Productivity Analysis 19: 211–226. Sickles, R. C., and B. Cigerli. 2009. “Krugman and Young Revisited: A Survey of the Sources of Productivity Growth in a World with Less Constraints.” Seoul Journal of Economics 22: 29–54. Sickles, R. C., and M. L. Streitwieser. 1998. “An Analysis of Technology, Productivity and Regulatory Distortion in the Interstate Natural Gas Transmission Industry: 1977–1985.” Journal of Applied Econometrics 13: 377–95. Sickles, R. C., and V. Zelenyuk. Forthcoming. Measurement of Productivity and Efficiency: Theory and Practice. New York: Cambridge University Press. Silva, E., A. O. Lansink and S. E. Stefanou. 2015. “The Adjustment-Cost Model of the Firm: Duality and Productive Efficiency.” International Journal of Production Economics 168: 245–256.

72 Emili Grifell-Tatjé, C. A. Knox Lovell, and Robin C. Sickles Silva, E., and S. E. Stefanou. 2003. “Nonparametric Dynamic Production Analysis and the Theory of Cost.” Journal of Productivity Analysis 19: 5–32. Silva, E., and S. E. Stefanou. 2007. “Dynamic Efficiency Measurement: Theory and Application.” American Journal of Agricultural Economics 89: 398–419. Sloan, A. P. 1964. My Years with General Motors. Garden City, NY: Doubleday. Smolny, W. 2000. “Sources of Productivity Growth: An Empirical Analysis with German Sectoral Data.” Applied Economics 32: 305–14. Solow, R. M. 1957. “Technical Change and the Aggregate Production Function.” Review of Economics and Statistics 39: 312–320. Solow, R. M. 1987. “We’d Better Watch Out.” New York Times, 36. July 12. Song, J., D. J. Price, F. Guvenen, N. Bloom, and T. von Wachter. 2015. “Firming Up Inequality.” National Bureau of Economic Research Working Paper 21199. http://www.nber.org/papers/ w21199. Squires, B. M. 1917. “Productivity and Cost of Labor in the Lumber Industry.” Monthly Labor Review 5: 66–79. Statistics Netherlands. 2015. ICT and Economic Growth. http://www.cbs.nl/en-GB/menu/ themas/macro-economie/publicaties/publicaties/archief/2015/2015-ict-and-economic- growth-pub.htm. Stewart, E. 1922. “Efficiency of American Labor.” Monthly Labor Review 15: 1–12. Stiglitz, J. E., A. Sen, and J-P Fitoussi. 2009. “The Measurement of Economic Performance and Social Progress Revisited.” OFCE Centre de recherché en économie de sciences Po. Working Paper 2009-33 (December). www.ofce.sciences-po.fr/pdf/dtravail/WP2009-33.pdf Syverson, C. 2011. “What Determines Productivity?” Journal of Economic Literature 49: 326–365. Tarbell, I. M. 1904. The History of the Standard Oil Company. New York: McClure, Phillips. Temin, P. (ed.). 1991. Inside the Business Enterprise: Historical Perspectives on the Use of Information. Chicago: University of Chicago Press. Timmer, M. P., A. A. Erumban, B. Los, R. Stehrer, and G. J. de Vries. 2014. “Slicing Up Global Value Chains.” Journal of Economic Perspectives 28: 99–118. Topp, V., and T. Kulys. 2014. “On Productivity: The Influence of Natural Resource Inputs.” International Productivity Monitor 27: 64–78. Topp, V., L. Soames, D. Parham and H. Bloch. 2008: “Productivity in the Mining Industry: Measurement and Interpretation.” Productivity Commission Staff Working Paper. www.pc.gov.au/research/staffworkingpaper/mining-productivity. Tulkens, H., and P. Vanden Eeckaut. 1995. “Nonparametric Efficiency, Progress and Regress Measures for Panel Data: Methodological Aspects.” European Journal of Operational Research 80: 474–499. van Ark, B. 2016. “The Productivity Paradox of the New Digital Economy.” International Productivity Monitor 31: 3–18. Van Reenen, J. 2011. “Does Competition Raise Productivity Through Improving Management Quality?” International Journal of Industrial Organization 29: 306–316. Vincent, A. L. A. 1968. La Mesure de la productivité. Paris: Dunod. Ward, B. 1958. “The Firm in Illyria: Market Syndicalism.” American Economic Review 48: 566–589. Winter, S. G., and G. Szulanski. 2001. “Replication as Strategy.” Organization Science 12: 730–743. Wolff, E. N. (2013), Productivity Convergence: Theory and Evidence. New York: Cambridge University Press.

Overview of Productivity Analysis 73 Wolff, F.-C., D. Squires, and P. Guillotreau. 2013. “The Firm’s Management in Production: Management, Firm, and Time Effects in an Indian Ocean Tuna Fishery.” American Journal of Agricultural Economics 95: 547–567. Young, A. 1992. “A Tale of Two Cities: Factor Accumulation and Technical Change in Hong Kong and Singapore.” NBER Macroeconomics Annual 1992: 13–54. Young, A. 1995. “The Tyranny of Numbers: Confronting the Statistical Realities of the East Asian Growth Experience.” Quarterly Journal of Economics 110: 641–680.

Pa rt I I

T H E F OU N DAT ION S OF P RODU C T I V I T Y A NA LYSI S

Chapter 2

Em pirical Produ c t i v i t y In di ces and I ndi c ators Bert M. Balk

2.1. Introduction Through its activities a firm creates value, and this has more dimensions than only the monetary.1 Accordingly, there is a multitude of perspectives coming from a multitude of stakeholders from which firm performance can be assessed (see Harrison and Wicks 2013). Marr (2012), for instance, discusses 75 measures, covering financial, customer, marketing and sales, operational processes and supply chain, employee, and corporate social responsibility perspectives. In this chapter, however, our attention will be restricted to economic performance measures such as profit, profitability, profit margin, productivity, price recovery, and their interlinkages. Though these concepts are primarily defined for actual economic agents such as plants or firms, they can easily be extended to industries, industrial sectors, or economies. This brings us to the realm of national accounts. What may the reader expect? In section 2.2 we begin by defining the key concepts: profit (revenue minus total cost), profitability (revenue divided by total cost), and two margins (profit-cost and profit-revenue). As the building blocks of these concepts are values, and we are interested in change through time, section 2.3 is devoted to the difference between direct and chained indices and indicators. The next question to consider is how to decompose value change into price and quantity components. Thus, section 2.4 provides a brief survey of index number theory, basically concentrating on price indices, the practice of deflation, and the practically important subject of two- stage indices. After all these preparations, section 2.5 turns to the heart of the matter, namely the definition of total factor productivity (TFP) and total price recovery indices as the two components of profitability change. Section 2.6 discusses the relation between our

78 Bert M. Balk definition and the practices of growth accounting and production function estimation. In the extended online version of this chapter, the closely related problem of the decomposition of a margin change is discussed. Section 2.7, finally, discusses the relation between total and partial productivity indices. Profit change is another key area. Since profit change is measured as a difference, indicators come into play—indicators being the difference-type counterparts from indices, which are ratio-type measures. Section 2.8 briefly reviews the theory, after which section 2.9—which can be considered as the correlate of section 2.5—defines TFP and total price recovery indicators. In the extended online version of this chapter, the fundamental equivalence of the two approaches, the multiplicative based on indices and the additive based on indicators, is demonstrated. In sections 2.2–2.9 the discussion is cast in terms of the so-called KLEMS-Y input– output model, where Y stands for physical output (goods and services). An important alternative is the KL-VA model, where value added serves as output. Value-added-based profitability change and its decomposition in price and quantity components is discussed in sections 2.10 and 2.11. For a discussion of the decomposition of value added change as such, the reader is referred to the extended online version of this chapter. Sections 2.2–2.11 basically consider the case of a single production unit moving through time. Section 2.12 continues by considering a dynamic ensemble of production units, and the relation between aggregate and subaggregate, or individual, measures of productivity change. The main results available in the literature are reviewed in two subsections, devoted to the top-down and the bottom-up approach, respectively. Section 2.13 recapitulates the main characteristics of the approach outlined in this chapter. Summarizing, this chapter shows what can be done with data alone, in particular (observable) nominal value data and (constructed) price index numbers (or, deflators).

2.2. Notation and Key Concepts We consider a single production unit through time and will later consider the comparison of multiple production units at a given point of time (a so-called cross- sectional or spatial comparison). This production unit may be a plant, a firm, an industry, an industrial sector, or even (the measurable part of) an entire economy. The unit is considered to be consolidated; that is, all within-unit transactions are netted out. At the output side of this unit we distinguish M items (goods and services), each with their unit price (as received by the production unit) pmt and quantity ymt, where m = 1, …, M , and t = 0, 1, …, T denotes an accounting period (say, a year). Similarly, at the input side we distinguish N items (goods, services, and assets), each with their unit price (as paid by the production unit) wnt and quantity xnt , where n = 1, …, N .

Empirical Productivity Indices and Indicators 79 Usually the input-side items are allocated to a few broadly defined classes:2 capital K , labor L, energy E, materials M, services S. Thus, we are primarily discussing the so- called KLEMS-Y input–output model.3 To avoid notational clutter, simple vector notation will be used throughout this chapter. All the prices and quantities are assumed to be positive, unless stated otherwise. The ex post accounting point of view will be used; that is, quantities and monetary values of the so-called flow variables (output Y , and LEMS inputs) are realized values, complete knowledge of which becomes available after the accounting period has expired. Similarly, the cost of capital input K is calculated ex post. This is consistent with official statistical practice. The unit’s revenue, that is, the value of its output, during the accounting period t is defined as

M

Rt ≡ pt ⋅ y t ≡ ∑ pmt ymt , (2.1) m =1

whereas its (total) production cost is defined as

N

C t ≡ w t ⋅ x t ≡ ∑ wnt xnt . (2.2) n =1

Given positive prices and quantities, it will always be the case that Rt > 0 and C t > 0. The unit’s profit (including taxes on production) is then given by its revenue minus its cost;4 that is,

Πt ≡ Rt − C t = pt ⋅ y t − w t ⋅ x t . (2.3)

The unit’s (gross-output based) profitability (also including taxes on production) is defined as its revenue divided by its cost; that is,

ϒt ≡ Rt /C t = pt ⋅ y t /w t ⋅ x t . (2.4)

Profit and profitability are similar financial performance concepts. The profitability concept seems to have been introduced by the economist Georgescu-Roegen in 1951. Whereas profit is a difference, profitability is a ratio. Profit can be positive, zero, or negative. Profitability is always positive, but can be greater than one, equal to one, or less than one. Since profitability is independent of the size of the production unit, for comparative purposes, profitability is a more natural measure than profit. The fundamental equivalence of the two concepts is expressed by the following relation:

Πt = LM (Rt , C t ) t ln ϒ

(Πt ≠ 0, ϒt ≠ 1), (2.5)

where LM (a, b) denotes the logarithmic mean.5

80 Bert M. Balk The next two performance concepts are margins. The first is the profit-cost margin of the production unit, defined as profit over cost, Πt . (2.6) Ct The relation between profit-cost margin and profitability is then given by

µt ≡

µt = ϒt − 1; (2.7)

that is, the profit-cost margin is the profitability expressed as a percentage (which is usually published as µt ×100%). But what precisely does this mean? To get a clue, consider first the single-output case; that is, M = 1. Then the production unit’s profitability reduces to

(

)

ϒt = pt y t /C t = pt / C t /y t ; (2.8)

that is, price over cost per unit. Put otherwise, the profit-cost margin µt is then the markup of price over unit cost. For the general, multi-output case, suppose that the cost can be allocated to the various outputs; that is, C t = Σ mM=1 Cmt, where Cmt is the cost for producing ymt units of output m (m = 1, , M ). Then the unit’s profitability can be decomposed as

∑ = ∑

M

ϒ

t

m =1 M

pmt ymt

m =1

Cmt

Cmt pmt . (2.9) t t t m = 1 C Cm / y m M

=∑

Thus profitability is a cost-share weighted mean of output-specific price over unit-cost relatives. Put otherwise, the profit-cost margin is a cost-share weighted mean of output- specific markups,

Cmt t µ m , (2.10) t m =1 C M

µt = ∑

where µtm ≡ pmt / (Cmt / ymt ) (m = 1, ..., M ). The second margin is the profit-revenue margin, defined as profit over revenue,6 Πt . (2.11) Rt This margin plays an important role in the duPont triangle, a management accounting system developed at duPont and General Motors early in the twentieth century (see Grifell-Tatjé and Lovell 2014). The relation between profit-revenue margin and profitability is given by

νt ≡

ν t = 1 − 1/ ϒt . (2.12)

Empirical Productivity Indices and Indicators 81 The profit-revenue margin is the percentage of revenue that is considered as profit (and usually published as ν t ×100%). The two margin concepts are connected by the profitability concept. Connecting the three definitions yields

ϒt = µt /ν t ; (2.13)

that is, profitability equals profit-cost margin divided by profit-revenue margin. Notice that the two margin concepts share with profit the property of being positive, zero, or negative.

2.3. Measuring Change As said, we are primarily concerned with the performance of our production unit through time. That is, we want to compare revenue, cost, profit, or profitability of a certain period t to an earlier period t ′ where t , t ′ = 0, 1, …, T . The key to the index number approach is that the availability of detailed data makes it possible to decompose any value change into two components: a price component and a quantity component. Now change can basically be measured in two ways: multiplicatively as a ratio and additively as a difference. The first leads to indices and the second to indicators. Let us consider indices first and take revenue as an example.7

2.3.1. Indices Let the periods t and t ′ be adjacent; that is, t ′ = t − 1. Given the price and quantity data and using the economic statistician’s toolkit, the ratio Rt /Rt −1 can be decomposed into two parts, Rt = P ( pt , y t , pt −1 , y t −1 )Q( pt , y t , pt −1 , y t −1 ), (2.14) Rt −1 where P(.) is a bilateral price index and Q(.) is a bilateral quantity index, the precise definitions of which will be postponed to the next section. When the periods t and t ′ are non-adjacent, that is, when t ′ = t − s with 1 < s ≤ t , then there are two options. The first is to use a set of bilateral indices to decompose the ratio Rt /Rt − s; that is,

Rt = P ( pt , y t , pt − s , y t − s )Q( pt , y t , pt − s , y t − s ). (2.15) t −s R

82 Bert M. Balk The second is to decompose the ratio Rt /Rt − s first as a multiplicative chain of adjacent- period ratios, t Rt Rs′ = ∏ s ′ −1 , (2.16) t −s R s ′= t − s +1 R

and then to decompose each of these adjacent-period ratios like expression (2.14), Rs′ = P ( p s ′ , y s ′ , p s ′ −1 , y s ′ −1 )Q( p s ′ , y s ′ , p s ′ −1 , y s ′ −1 ) s ′ = t − s + 1, …, t . (2.17) s ′ −1 R Substituting expression (2.17) into (2.16) and rearranging delivers as an alternative decomposition of the ratio Rt /Rt − s the more complicated expression

(

)

t t Rt = ∏ P ( p s ′ , y s ′ , p s ′ −1 , y s ′ −1 ) × ∏ Q( p s ′ , y s ′ , p s ′ −1 , y s ′ −1 ). (2.18) t −s R s ′= t − s +1 s ′= t − s +1 The first factor at the right-hand side of this expression is called a chained price index, and the second factor is called a chained quantity index. It is important to observe that in general, on the same data, the decompositions (2.15) and (2.18) deliver different outcomes. The first decomposition is in terms of direct indices. A direct index is a bilateral index; that is, a function depending only on the data of the two periods t and t − s . The second decomposition is in terms of chained indices. A chained index is a multilateral index; that is, a function that does not only depend on the data of the two periods being compared, t and t − s respectively, but also on the data of all the intervening periods.

2.3.2. Indicators Let the periods t and t ′ be adjacent; that is, t ′ = t − 1. Given the price and quantity data and using the economic statistician’s toolkit, the difference Rt − Rt −1 can be decomposed into two parts,

Rt − Rt −1 = ( pt , y t , pt −1 , y t −1 ) + ( pt , y t , pt −1 , y t −1 ), (2.19)

where (.) is a bilateral price indicator and (.) is a bilateral quantity indicator, the precise definitions of which are provided in a later section of the chapter. Of course, a difference such as Rt − Rt −1 only makes sense when the two money amounts involved are deflated by some general inflation measure (such as the headline Consumer Price Index, or CPI), or when the two periods are so close that deflation is deemed unnecessary. In the remainder of this subsection, when discussing difference measures, either of the two situations is tacitly presupposed.

Empirical Productivity Indices and Indicators 83 When the periods t and t ′ are non-adjacent, that is, when t ′ = t − s with 1 < s ≤ t, then there are two options. The first is to use a set of bilateral indicators to decompose the difference Rt − Rt − s ; that is,

Rt − Rt − s = ( pt , y t , pt − s , y t − s ) + ( pt , y t , pt − s , y t − s ). (2.20)

The second is to decompose the difference Rt − Rt − s first as an additive chain of adjacent- period differences, Rt − Rt − s =

t

∑

(R s ′ − R s ′ −1 ), (2.21)

s ′= t − s +1

and then to decompose each of these adjacent-period differences like expression (2.19), R s ′ − R s ′ −1 = ( p s ′ , y s ′ , p s ′ −1 , y s ′ −1 ) + ( p s ′ , y s ′ , p s ′ −1 , y s ′ −1 )

(s ′ = t − s + 1, ..., t ). (2.22)

Substituting expression (2.22) into (2.21) and rearranging delivers as an alternative decomposition of the difference Rt − Rt − s the more complicated expression

Rt − Rt − s =

t

∑

s ′= t − s +1

( p s ′ , y s ′ , p s ′ −1 , y s ′ −1 ) +

t

∑

( p s ′ , y s ′ , p s ′ −1 , y s ′ −1 ). (2.23)

s ′= t − s +1

The first factor at the right-hand side of this expression is called a chained price indicator, and the second factor is called a chained quantity indicator. It is important to observe that in general, on the same data, the decompositions (2.20) and (2.23) deliver different outcomes. The first decomposition is in terms of direct indicators. A direct indicator is a bilateral function, depending only on the data of the two periods t and t − s . The second decomposition is in terms of chained indicators. A chained indicator is a multilateral function, depending on the data of the two periods being compared, t and t − s respectively, and on the data of all the intervening periods.

2.3.3. Conclusion From the preceding two subsections, it is clear that there is no loss of generality when in the remainder of this survey we consider only the comparison of adjacent periods. To simplify notation, the two periods we consider are t = 1 (which will be called the comparison period) and t ′ = 0 (which will be called the base period).

2.4. Indices In this section we consider an aggregate consisting of M items (goods and services) with (unit) prices going from pm0 to pm1 and quantities going from ym0 to y1m (m = 1, , M ). The

84 Bert M. Balk first subsection reviews price indices, the second subsection reviews quantity indices, and the third subsection reviews the concept of two-stage indices. It is important to note that, though stated in terms of output, the theory surveyed in this section is equally applicable to other situations after appropriate modification of the number of items and the interpretation of the variables.

2.4.1. Price Indices Formally, a bilateral price index is a positive, continuously differentiable function P ( p1 , y1 , p0 , y 0 ) : ℜ4+M+ → ℜ+ + that, except in extreme situations, correctly indicates any increase or decrease of the elements of the price vectors p1 or p0, conditional on the quantity vectors y1 and y 0. Likewise, a bilateral quantity index is a positive, continuously differentiable function of the same variables Q( p1 , y1 , p0 , y 0 ) : ℜ4+M+ → ℜ+ + that, except in extreme situations, correctly indicates any increase or decrease of the elements of the quantity vectors y1 or y 0, conditional on the price vectors p1 and p0. The number M is called the dimension of the price or quantity index. These two definitions are deliberately kept pretty vague, to accommodate official and unofficial practice. Over the last century, index number theory has delivered much more precise statements in the form of axioms and/or tests, but it was soon discovered that no index satisfies simultaneously all those requirements. The basic requirements on price and quantity indices comprise the following: 1. that they exhibit the desired monotonicity properties globally; that is, P ( p1 , y1 , p0 , y 0 ) is nondecreasing in p1 and nonincreasing in p0, and Q( p1 , y1 , p0 , y 0 ) is nondecreasing in y1 and nonincreasing in y 0 ; 2. that P ( p1 , y1 , p0 , y 0 ) is linearly homogeneous in p1 and Q( p1 , y1 , p0 , y 0 ) is linearly homogeneous in y1 ; 3. that they satisfy the Identity Test; that is, P ( p0 , y1 , p0 , y 0 ) = 1 and Q( p1 , y 0 , p0 , y 0 ) = 1 ; 4. that P ( p1 , y1 , p0 , y 0 ) is homogeneous of degree 0 in ( p1 , p0 ) , and Q( p1 , y1 , p0 , y 0 ) is homogeneous of degree 0 in ( y1 , y 0 ) ; 5. that they are invariant to changes in the units of measurement of the items. The reader is referred to Balk (2008, 56–61) for the formalization and an extensive discussion of these and other requirements. Though P ( p1 , y1 , p0 , y 0 ) and Q( p1 , y1 , p0 , y 0 ) are both defined as functions of 4M variables, it appears that any such function that is invariant to changes in the units of measurement can be written as a function of only 3M variables, namely the price relatives pm1 / pm0 or the quantity relatives y1m / ym0 , the comparison period values v1m ≡ pm1 y1m, and the base period values vm0 ≡ pm0 ym0 (m = 1, , M ). This feature appears to be very useful

Empirical Productivity Indices and Indicators 85 in practice, especially when M is large. Price and quantity indices can then be estimated from samples of items for which price or quantity relatives are observed and for which population value information is available. We review the most important functional forms, starting with price indices. The Laspeyres price index is conventionally defined as a function of prices and quantities, P L ( p1 , y1 , p0 , y 0 ) ≡ p1 ⋅ y 0 / p0 ⋅ y 0 ,

but notice that this index can be written as a function of price relatives and (base period) values or value shares, M

P L ( p1 , y1 , p0 , y 0 ) = ∑ ( pm1 /pm0 ) vm0

0 m

M

m =1

m =1

∑ vm0 = ∑ sm0 ( pm1 /pm0 ),

m =1

0 m

M

M 0 m =1 m

where s ≡ v /Σ v (m = 1, , M ) are base period value shares. Put otherwise, the Laspeyres price index is a weighted arithmetic mean of price relatives, using base period value shares as weights. Similarly, the Paasche price index, conventionally defined as P P ( p1 , y1 , p0 , y 0 ) ≡ p1 ⋅ y1 /p0 ⋅ y1 ,

can be written as a function of price relatives and (comparison period) values or value shares,

M P ( p , y , p , y ) = ∑ ( pm0 /pm1 ) v1m m =1 P

1

1

0

0

v ∑ m =1 M

1 m

−1

−1

M = ∑ s1m ( pm1 /pm0 )−1 , m =1

where s1m ≡ v1m /Σ mM=1 v1m (m = 1, , M ) are comparison period value shares. Thus the Paasche price index is a weighted harmonic mean of price relatives, using comparison period value shares as weights. Both indices are by definition asymmetric: the Laspeyres index compares prices by means of base-period quantities, whereas the Paasche index does this by means of comparison-period quantities. A symmetric alternative is provided by the Marshall- Edgeworth index, which uses the (arithmetic) mean of base and comparison period quantities as instrument for comparing prices; that is,

P ME ( p1 , y1 , p0 , y 0 ) ≡ p1 ⋅ ( y 0 + y1 ) / p0 ⋅ ( y 0 + y1 ).

A little bit of manipulation reveals that this index can be written as a function of values, Laspeyres, and Paasche indices,

P ME ( p1 , y1 , p0 , y 0 ) =

p0 ⋅ y 0 P L ( p1 , y1 , p0 , y 0 ) + p1 ⋅ y1 p0 ⋅ y 0 + p1 ⋅ y1 / P P ( p1 , y1 , p0 , y 0 )

.

86 Bert M. Balk A different symmetric index is the Fisher price index, defined as the geometric mean of the Laspeyres and Paasche indices:

P F ( p1 , y1 , p0 , y 0 ) ≡ P L (.)P P (.)

1/ 2

Σ M s 0 ( p1 / p0 ) = Mm =11 m 1 m 0 m −1 . Σ m =1sm ( pm / pm )

1/ 2

The Geometric Laspeyres price index is defined as the base period value-share weighted geometric mean of the price relatives,

M

P GL ( p1 , y1 , p0 , y 0 ) ≡ ∏ ( pm1 / pm0 )sm . 0

m =1

GL

It is a mathematical fact that P ( p , y , p0 , y 0 ) ≤ P L ( p1 , y1 , p0 , y 0 ) and that the difference between these two index numbers depends on the variance of the price relatives pm1 / pm0 (m = 1, , M ). The Geometric Paasche price index is the alternative, using comparison period value shares as weights,

1

1

M

P GP ( p1 , y1 , p0 , y 0 ) ≡ ∏ ( pm1 / pm0 )sm . 1

m =1

Due to the same mathematical theorem, it is the case that P GP ( p1 , y1 , p0 , y 0 ) ≥ P P ( p1 , y1 , p0 , y 0 ). Notice the difference between this and the previous inequality. The geometric mean of the Geometric Laspeyres and Paasche price indices is known as the Törnqvist price index,

P T ( p1 , y1 , p0 , y 0 ) ≡ P GL (.)P GP (.)

1/ 2

M

(s

= ∏ ( pm1 / pm0 )

0 1 m + sm

)/2 .

m =11

Since the Marshall- Edgeworth, Fisher, and Törnqvist price indices are symmetric with respect to weights, they satisfy the Time Reversal Test requiring that P ( p0 , y 0 , p1 , y1 ) = 1 / P ( p1 , y1 , p0 , y 0 ).8 Whether, at the same data, P T (.) ≥ P F (.) or P T (.) ≤ P F (.) is an empirical question. As Diewert (1978, Theorem 5) has shown, at any point where p1 = p0 and y1 = y 0 the indices P T (.) and P F (.) differentially approximate each other to the second order. To facilitate a comparison of Fisher and Törnqvist price indices, Balk (2008, section 4.2.3) showed that the Fisher price index can be expressed in geometric mean form as M s 0 LM ( pm1 / pm0 , P L (.)) ln P F ( p1 , y1 , p0 , y 0 ) = (1 / 2)∑ M m 0 1 L 0 m =1 Σ m =1 sm LM ( pm / pm , P (.)) (2.24) sm01LM ( pm1 / pm0 , P P (.)) 1 0 + M 01 ln( pm / pm ), Σ m =1sm LM ( pm1 / pm0 , P P (.))

Empirical Productivity Indices and Indicators 87 where sm01 ≡ pm0 y1m / p0 ⋅ y1 (m = 1,..., M ). This then can be compared to a similar expression for the Törnqvist price index, M

ln P T ( p1 , y1 , p0 , y 0 ) = (1 / 2)∑[sm0 + sm1 ]ln( pm1 / pm0 ). (2.25)

m =1

It is clear that the difference between the two indices depends on the variance of the price relatives pm1 / pm0 (m = 1, …, M ).

2.4.2. Deflation Rather than providing a similarly structured review of the main quantity indices, we will pursue what happens when a value ratio is deflated by a price index—that is, when the quantity index is defined by Q( p1 , y1 , p0 , y 0 ) ≡

p1 ⋅ y1 P ( p1 , y1 , p0 , y 0 ). (2.26) p0 ⋅ y 0

Notice that this is equivalent to stating that P(.) and Q(.) satisfy the Product Test. Let us review the earlier examples. As one checks easily, deflating a value ratio by a Laspeyres price index delivers a Paasche quantity index, −1

M Q P ( p1 , y1 , p0 , y 0 ) ≡ p1 ⋅ y1 / p1 ⋅ y 0 = ∑ s1m ( y1m / ym0 )−1 . m =1

Similarly, deflating a value ratio by a Paasche price index delivers a Laspeyres quantity index, M

Q L ( p1 , y1 , p0 , y 0 ) ≡ p0 ⋅ y1 / p0 ⋅ y 0 = ∑ sm0 ( ym1 / ym0 ).

m =1

It follows then quickly that deflating a value ratio by a Fisher price index delivers a Fisher quantity index,

Q ( p , y , p , y ) ≡ Q (.)Q (.) F

1

1

0

0

L

P

1/ 2

1/ 2

Σ M s 0 ( y1 / y 0 ) = Mm =11 m 1 m 0 m −1 . Σ m =1 sm ( y m / y m )

Fisher indices are examples of so-called ideal indices; that is, they satisfy the Product Test and have the same functional form (after interchanging prices and quantities, of course).

88 Bert M. Balk Deflating a value ratio by a Marshall-Edgeworth price index delivers 1 + Q L ( p1 , y1 , p0 , y 0 ) ; 1 + 1 / Q P ( p1 , y1 , p0 , y 0 )

an expression that is not linearly homogeneous in comparison period quantities (see also Balk 2008, 83). Deflating a value ratio by a Geometric Laspeyres price index delivers an expression that cannot be simplified: p1 ⋅ y1 p0 ⋅ y 0

M

∏(p m =1

1 m

0

/ pm0 )sm .

As one easily checks, this expression does not satisfy one of the fundamental requirements for a quantity index, namely that such an index satisfies the Identity Test. Similarly, deflating a value ratio by a Geometric Paasche price index delivers an expression that does not satisfy the Identity Test, and a fortiori the same holds when a value ratio is deflated by a Törnqvist price index. In particular, it is important to notice that the Törnqvist indices are not ideal; that is,

p1 ⋅ y1 p0 ⋅ y 0

M

P T ( p1 , y1 , p0 , y 0 ) ≠ QT ( p1 , y1 , p0 , y 0 ) ≡ ∏ ( ym1 / ym0 )( sm + sm )/2 , 0

1

m =1

though in practice the discrepancy often may be negligible. Notice that the Törnqvist quantity index QT ( p1 , y1 , p0 , y 0 ) does satisfy the Identity Test. Do there exist price and quantity indices of the geometric mean form that are ideal? The answer appears to be positive. Sato-Vartia indices are defined as M

P SV ( p1 , y1 , p0 , y 0 ) ≡ ∏ ( pm1 / pm0 )φm

m =1 M

Q SV ( p1 , y1 , p0 , y 0 ) ≡ ∏ ( ym1 / ym0 )φm ,

m =1

where φm ≡ LM (s , s ) Σ LM (s , s ) (m = 1, …, M ). The Sato- Vartia indices are, like the Törnqvist indices, geometric means. The difference is in the weights: a Törnqvist index weighs the individual price or quantity relatives with arithmetic mean value shares, whereas a Sato- Vartia index weighs with (normalized) logarithmic mean shares. It is a straightforward exercise to check that P SV ( p1 , y1 , p0 , y 0 ) × Q SV ( p1 , y1 , p0 , y 0 ) = p1 ⋅ y1 / p0 ⋅ y 0 . 0 m

1 m

M m =1

0 m

1 m

2.4.3. Two-Stage Indices The number of items distinguished in a particular aggregate (M) may be very large. To accommodate this, (detailed) classifications are used, by which all the items are allocated to hierarchically organized (sub-)aggregates. The calculation of price and/or

Empirical Productivity Indices and Indicators 89 quantity indices then proceeds in stages. Theoretically, it suffices to distinguish only two stages. At the first stage, one calculates indices for all the subaggregates at a certain level; at the second stage, these subaggregate indices are combined to aggregate indices. Let us be more precise. Let the aggregate under consideration, comprising all the M items, be partitioned arbitrarily into K disjunct subaggregates Ak , A ≡ {1, …, M } = ∪kK=1 Ak , Ak ∩ Ak ′ = ∅

(k ≠ k ′).

Each subaggregate consists of a number of items. Let M k ≥ 1 denote the number of items contained in Ak (k = 1, , K ). Obviously M = Σ kK=1 M k. Let ( p1k , y1k , pk0 , y k0 ) be the subvector of ( p1 , y1 , p0 , y 0 ) corresponding to the subaggregate Ak . Recall that vmt ≡ pmt ymt is the value of item m at period t. Then Vkt ≡ Σ m∈Ak vmt (k = 1, …, K ) is the value of subaggregate Ak at period t, and V t ≡ Σ m∈A vmt = Σ kK=1Vkt is the value of the aggregate A at period t. Let P (.), P (1) (.), …, P ( K ) (.) be price indices of dimension K , M1 ,..., M K , respectively, that satisfy the basic requirements. Recall that any such price index can be written as a function of price relatives and base and comparison period item values. Now we replace in P(.) the price relatives by subaggregate price indices and the item values by subaggregate values. Then the price index defined by P * ( p1 , y1 , p0 , y 0 ) ≡ P (P ( k ) ( p1k , y1k , pk0 , y k0 ), Vk1 , Vk0 ; k = 1, ..., K ) (2.27)

is of dimension M and also satisfies the basic requirements. The index P * (.) is called a two-stage price index. The first stage refers to the indices P (k ) (.) for the subaggregates Ak (k = 1, ..., K ). The second stage refers to the index P(.) that is applied to the subindices P (k ) (.) (k = 1, ..., K ). A two-stage index such as defined by expression (2.27) closely corresponds to the calculation practice at statistical agencies. All the subindices are then usually of the same functional form, for instance, Laspeyres or Paasche indices. The aggregate, second-stage index may or may not be of the same functional form. This could be, for instance, a Fisher index. If the functional forms of the subindices P (k ) (.) (k = 1, , K ) and the aggregate index P(.) are the same, then P * (.) is called a two-stage P(.)-index. Continuing our example, the two-stage Laspeyres price index reads K

P * L ( p1 , y1 , p0 , y 0 ) ≡ ∑ P L ( p1k , y1k , pk0 , y k0 )Vk0

k =1

K

∑V k =1

0 k

,

and, by inserting the various definitions, one easily checks that the two-stage and the single-stage Laspeyres price indices coincide:

K

P (p , y , p , y ) = ∑ *L

1

1

0

0

k =1

K 1 0 p1k ⋅ y k0 Vk0 Σ k =1Σ m∈Ak pm ym = = P L ( p1 , y1 , p0 , y 0 ). Vk0 V 0 V0

90 Bert M. Balk This, however, is the exception rather than the rule. For most indices, two-stage and single-stage variants do not coincide. Indices for which two-stage and single-stage variants coincide are called consistent in aggregation (CIA). Though the Fisher price index is not CIA, fortunately single-stage and two-stage Fisher indices appear to be close approximations of each other (as shown by Diewert 1978, Appendix 2). Two-stage quantity indices are defined similarly. Thus, let Q(.), Q (1) (.), ..., Q ( K ) (.) be quantity indices of dimension K , M1 ,..., M K , respectively, that satisfy the basic requirements. Then the quantity index defined by

Q * ( p1 , y1 , p0 , y 0 ) ≡ Q(Q ( k ) ( p1k , y1k , pk0 , y k0 ), Vk1 , Vk0 ; k = 1, ..., K ) (2.28)

is of dimension M and also satisfies the basic requirements. The index Q * (.) is called a two-stage quantity index. If the functional forms of the subindices Q (k ) (.) (k = 1, ..., K ) and the aggregate index Q(.) are the same, then Q * (.) is called a two-stage Q(.)-index. It is important to observe that not any two-stage construct satisfies the basic requirements, even if the constituent parts do. Consider, for instance,

(

)

QT (Vk1 / Vk0 ) / P T (k ) ( p1k , y1k , pk0 , y k0 ), Vk1 , Vk0 ; k = 1, ..., K .

At the first stage, subaggregate quantity indices are calculated as value ratios deflated by Törnqvist price indices. At the second stage, these quantity indices are aggregated by means of the Törnqvist quantity index. It is a straightforward exercise to check that the entire construct does not satisfy the Identity Test.

2.5. Decomposing a Profitability Ratio We now return to our main topic. The development over time of profitability is, rather naturally, measured by the ratio ϒ1 / ϒ 0. How to decompose this into a price and a quantity component? By noticing that ϒ1 R1 / C1 R1 / R0 = = (2.29) ϒ 0 R 0 / C 0 C1 / C 0 we see that the question reduces to the question of how to decompose the revenue ratio R1 / R0 and the cost ratio C1 / C 0 into two parts. The revenue ratio has been discussed extensively in the previous section. Thus there exist price and quantity indices P(.) and Q(.) such that

R1 = P ( p1 , y1 , p0 , y 0 ) Q( p1 , y1 , p0 , y 0 ) 0 (2.30) R ≡ PR (1, 0) QR (1, 0),

Empirical Productivity Indices and Indicators 91 where the second line serves to define our shorthand notation. The cost ratio is structurally the same as the revenue ratio. Thus there exist also price and quantity indices P(.) and Q(.) such that the cost ratio can be decomposed as

C1 = P (w1 , x1 , w 0 , x 0 ) Q(w1 , x1 , w 0 , x 0 ) (2.31) C0 ≡ PC (1, 0)QC (1, 0).

All these price and quantity indices may or may not exhibit the same functional form. Ideally, they all satisfy the basic requirements. Notice that the dimensions of the indices in expressions (2.30) and (2.31) will usually be different. The subscripts R and C are used because, as will appear later, there are more output and input concepts. Using the two relations (2.30) and (2.31), the profitability ratio can be decomposed as

ϒ1 PR (1, 0) QR (1, 0) . (2.32) = ϒ 0 PC (1, 0) QC (1, 0)

This is the fundamental equation of the KLEMS-Y model, and basic for much of what follows. The TFP index, for period 1 relative to period 0, is now defined by

ITFPROD(1, 0) ≡

QR (1, 0) . (2.33) QC (1, 0)

Thus ITFPROD(1, 0) is the real or quantity component of the profitability ratio. Put otherwise, it is the ratio of an output quantity index and an input quantity index; ITFPROD(1, 0) is the factor with which the output quantities on average have changed relative to the factor with which the input quantities on average have changed. If the ratio of these factors is larger (smaller) than 1, there is said to be productivity increase (decrease).9 Notice that ITFPROD(1, 0) is a function of all the output and input prices and quantities; that is ITFPROD(1, 0) = ITFPROD( p1 , y1 , w1 , x1 , p0 , y 0 , w 0 , x 0 ). Its properties follow from those of the output and input quantity indices Q( p1 , y1 , p0 , y 0 ) and Q(w1 , x1 , w 0 , x 0 ). In particular, • in the single-input–single-output case (N = M = 1) the TFP index reduces to y1 / y 0 y1 / x1 ITFPROD(1, 0) = 1 0 = 0 0 ; that is, the ratio of the productivities in the two x /x y /x periods compared; • ITFPROD(1, 0) exhibits the desired monotonicity properties: nondecreasing in y1 , nonincreasing in x1 , nonincreasing in y 0 , nondecreasing in x 0 ;

92 Bert M. Balk • ITFPROD(1, 0) exhibits proportionality in input and output quantities; that is, ITFPROD( p1 , µ y 0 , w1 , λ x 0 , p0 , y 0 , w 0 , x 0 ) = µ / λ (λ, µ > 0) . Notice also that, again using expressions (2.30) and (2.31), there appear to be three other, equivalent representations of the TFP index, namely

ITFPROD(1, 0) =

=

(R1 / R0 ) / PR (1, 0) (2.34) (C1 / C 0 ) / PC (1, 0)

(R1 / R 0 ) / PR (1, 0) (2.35) QC (1, 0) Q (1, 0) . (2.36) = 1 R0 (C / C ) / PC (1, 0)

Put in words, we are seeing here, respectively, a deflated revenue index divided by a deflated cost index, a deflated revenue index divided by an input quantity index, and an output quantity index divided by a deflated cost index. The other part of the profitability ratio decomposition (2.32) is called total price recovery (TPR) index:

ITPR(1, 0) ≡

PR (1, 0) . (2.37) PC (1, 0)

This index measures the extent to which input price change is recovered by output price change. Now, if revenue change equals cost change, R1 / R0 = C1 / C 0 (for which zero profit in the two periods is a sufficient condition), then it follows that

ITFPROD(1, 0) = 1 / ITPR(1, 0); (2.38)

that is, the TFP index is equal to the inverse of the TPR index. Thus, zero profit does not imply zero productivity growth; it only implies that productivity growth is compensated by price recovery. In general, however, the dual productivity index 1 / ITPR(1, 0) will differ from the primal one, ITFPROD(1, 0). A non-market unit is characterized by the fact that there is no revenue, which limits the usefulness of expressions (2.32)–(2.36). But if there is some prices-free output quantity index Q( y1 , y 0 ), then the TFP index, for period 1 relative to period 0, is naturally defined by Q( y1 , y 0 ) / QC (1, 0). An alternative expression is obtained by replacing the input quantity index by the deflated cost index, Q( y1 , y 0 ) /[(C1 / C 0 ) / PC (1, 0)]. In the single- output case (that is, M = 1) this reduces to an index of unit cost divided by an input price index, [( y1 / C1 ) / ( y 0 / C 0 )]/ PC (1, 0)].

Empirical Productivity Indices and Indicators 93

2.6. Growth Accounting and Production Functions 2.6.1. Growth Accounting The foregoing definitions are already sufficient to provide examples of simple but useful analysis. First, reconsider relation (2.35), and rewrite this as

R1 / R0 = ITFPROD(1, 0) × QC (1, 0) × PR (1, 0). (2.39)

Taking logarithms, one obtains

ln(R1 / R0 ) = ln ITFPROD(1, 0) + ln QC (1, 0) + ln PR (1, 0). (2.40)

This relation, implemented with Fisher and Törnqvist indices, was used by Dumagan and Ball (2009) for an analysis of the US agricultural sector. Further, revenue change through time is only interesting insofar as it differs from general inflation. Hence, it makes sense to deflate the revenue ratio, R1 / R0, by a general inflation measure such as the (headline) CPI. Doing this, the previous equation can be written as

R1 / R 0 PR (1, 0) = ln ITFPROD(1, 0) + ln QC (1, 0) + ln . (2.41) ln 1 0 CPI 1 / CPI 0 CPI / CPI

Lawrence, Diewert, and Fox (2006) basically used this relation to decompose “real” revenue change into three factors: productivity change, input quantity change (which can be interpreted as measuring change of the production unit’s size), and “real” output price change, respectively. Our second example follows from rearranging expression (2.36) and taking logarithms. This delivers the following relation:

ln(C1 / C 0 ) = ln PC (1, 0) + ln QR (1, 0) − ln ITFPROD(1, 0). (2.42)

This relation was also used by Dumagan and Ball (2009).10 A further rearrangement gives

C1 / C 0 ln = ln PC (1, 0) − ln ITFPROD(1, 0). (2.43) QR (1, 0)

We see here that the growth rate of average cost can be decomposed into two factors, namely the growth rate of input prices and a residual that is the negative of TFP growth. Put otherwise, in the case of stable input prices, the growth rate of average cost is equal to minus the TFP growth rate.

94 Bert M. Balk Our third example follows from rearranging expression (2.34) as

ITFPROD(1, 0) =

1 + µ1 1 PC (1, 0), (2.44) 1 + µ 0 PR (1, 0)

where the profit-cost margin µt was defined by expression (2.6). Expression (2.44) can be read as picturing the distribution of the fruits of productivity increase: to the owner(s) of the production unit as the profit-cost margin increase, or to the customers as output prices decrease, or to the suppliers (including employees) as input prices (including wages) increase. Taking logarithms and rearranging a bit delivers the following relation:

1 + µ1 + ln PC (1, 0) − ln ITFPROD(1, 0), (2.45) ln PR (1, 0) = ln 1 + µ 0

This relation analyses output price change as resulting from three factors: change of the profit-cost margin, input price change, and, with a negative sign, TFP change. This relation may be used for regulatory purposes. For example, if the profit-cost margin is not allowed to grow, then output prices may on average rise as input prices do, but adjusted by the percentage of TFP growth. All these are examples of what is called growth accounting. The relation between index number techniques and growth accounting techniques can, more generally, be seen as follows. Recall the generic definition of the TFP index in expression (2.33), and rewrite this expression as follows:

QR (1, 0) = ITFPROD(1, 0) × QC (1, 0). (2.46)

Taking logarithms, this multiplicative expression can be rewritten as an additive one,

ln QR (1, 0) = ln ITFPROD(1, 0) + ln QC (1, 0). (2.47)

For index numbers in the neighborhood of 1, the logarithms thereof approximate percentages, and the last expression can thus be interpreted as saying that the percentage change of output volume equals the percentage change of input volume plus the percentage change of TFP. Growth accounting economists like to work with equations expressing output volume growth in terms of input volume growth plus a residual that is interpreted as TFP growth, thereby suggesting that the last two factors cause the first. However, TFP change cannot be considered as an independent factor since it is defined as output quantity change minus input quantity change. Put otherwise, a growth accounting table is nothing but an alternative way of presenting TFP growth and its contributing factors. And decomposition does not at all imply something about causality.

Empirical Productivity Indices and Indicators 95

2.6.2. Production Functions The identity in expression (2.46) is formulated in terms of indices. Economists, however, usually prefer working with levels. To transform the identity, we multiply both sides by base period revenue R0 ,

R0QR (1, 0) = ITFPROD(1, 0) × (R0 /C 0 ) × C 0QC (1, 0). (2.48)

Now the Product Test, expression (2.26), tells us that base period value times quantity index equals comparion period value divided by (deflated by) price index. Thus expression (2.48) can be rewritten as

R1 / PR (1, 0) = ITFPROD(1, 0) × (R0 / C 0 ) × C1 /PC (1, 0). (2.49)

These two equations are identities with two degrees of freedom each. Choosing particular functional forms for the price indices PR (1, 0) and PC (1, 0), or the quantity indices QR (1, 0) and QC (1, 0), fixes the outcome of the TFP index ITFPROD(1, 0). Consider as a special case the input quantity index

QC (1, 0) ≡ ∏ Qk (1, 0)αk where all α k > 0 and k ∈

∑α

k ∈

k

= 1. (2.50)

This is a two-stage quantity index. The first stage consists of subindices Qk (1, 0) for the input classes k ∈ ≡ {K , L, E , M , S}. The second stage combines (aggregates) these subindices with help of a (generalized) Cobb-Douglas function. Substituting expression (2.50) into expression (2.48) delivers

R0QR (1, 0) = ITFPROD(1, 0) ×

R0 C0 × × ∏ (Ck0Qk (1, 0))αk , (2.51) C 0 Π k ∈ (Ck0 )αk k ∈

which, again due to the Product Test, can also be written as

R1 / PR (1, 0) = ITFPROD(1, 0) ×

R0 C0 × × ∏ (Ck1 / Pk (1, 0))αk . (2.52) C 0 Π k ∈ (Ck0 )αk k ∈

Now this is beginning to look familiar. The left-hand side of this equation is called real output; that is, deflated revenue. Similarly, Ck1 /Pk (1, 0) is real input (that is, deflated cost) of class k ∈. Given suitable price indices, these are all measurable magnitudes. Next, ITFPROD(1, 0) is interpreted as technological change. Recall that R0 / C 0 = ϒ 0 = 1 + µ 0. If real output can be considered as a single output, then µ 0 is the base period markup. Finally, the term C 0 /Π k ∈ (Ck0 )αk is a remainder. Actually, it is the reciprocal of the weighted geometric mean of the cost shares of the input classes Ck0 / C 0 (k ∈) (recall that C 0 = Σ k ∈Ck0).

96 Bert M. Balk Thus, summing up, for any production unit, real output is related to aggregate real input, TFP change, and a constant covering base period markup and nonlinearity. The coefficients α k of the Cobb-Douglas function are as yet unspecified. Specifying them would immediately determine TFP change. Now what basically happens in production- function estimation is that these coefficients are estimated from a sample of production units. The procedure can be stylized as follows in a number of steps: • For a (balanced) panel of production units, say from a specific industry, real output and real inputs are given (whereby the use of specific deflators is presupposed). • TFP change is seen as consisting of two components, one industry-specific and the other production-unit-specific (idiosyncratic); the second component is usually considered as (random) noise. • The coefficients α k are assumed to be the same for all the production units and must be estimated. Transforming expression (2.52) into logarithmic form, we obtain

ln(R1 / PR (1, 0)) = ln ITFPROD(1, 0) + ln α 0 + ∑ α k ln(Ck1 / Pk (1, 0)), (2.53) k ∈

0 αk k

where α 0 ≡ R /Π k ∈ (C ) . This is the typical form of a Cobb-Douglas “production function.”11 Notice that this form deviates from that usually shown in textbooks and research articles. Conventionally, the Cobb-Douglas production function, like any other such function, is stated in terms of quantities. However, in most if not all realistic multiple-input multiple-output settings, quantities of inputs and outputs are not given. The best one can hope for is obtaining good proxies for real values. Put otherwise, there are only data to estimate equations such as given by expression (2.53).12 This and the previous expressions in a way condense all the outputs, the number M of which can be very large, into a single variable. Suppose that we are interested in a special set of outputs, say A, and let B denote the remainder. Let the aggregate output quantity index QR (1, 0) be a two-stage index of the Cobb-Douglas form, QR (1, 0) = QA (1, 0)β QB (1, 0)1−β, where 0 < β < 1 and QA (1, 0) and QB (1, 0) are suitable quantity indices for the subaggregates A and B, respectively. Then 0

R0QR (1, 0) =

R0 (RA0 QA (1, 0))β (RB0 QB (1, 0))1−β (2.54) 0 1− β ( R ) ( RB ) 0 β A

or, due to the Product Test,

R1 / PR (1, 0) =

R0 (R1A / PA (1, 0))β (R1B / PB (1, 0))1−β , (2.55) (RA0 )β (RB0 )1−β

Empirical Productivity Indices and Indicators 97 or, in logarithms,

ln(R1 / PR (1, 0)) = ln γ + β ln(R1A / PA (1, 0)) + (1 − β)ln(R1B / PB (1, 0)), (2.56)

where γ ≡ R0 / (RA0 )β (RB0 )1−β. Substituting now expression (2.56) into expression (2.53), and rearranging a bit, delivers ln(R1A / PA (1, 0)) = (1 / β)ln ITFPROD(1, 0) + (1 / β)ln(α 0 / γ ) (2.57)

+ ∑ (α k / β)ln(Ck1 / Pk (1, 0)) − ((1 − β) / β)ln(R1B / PB (1, 0)) k ∈

Here we see the logarithm of real output of subaggregate A expressed as a linear function of TFP change, a constant, and a weighted sum of all the logarithmic real inputs minus the logarithmic real output of the remainder subaggregate B. Expression (2.57) corresponds to the “production function” estimated by Dhyne et al. (2014). Notice, however, that the parameters in this expression are interrelated, unlike the parameters in expression (2.53).

2.7. Partial Productivity Measures The distinguishing feature of a TFP index, such as in the KLEMS- Y model ITFPROD(1, 0), is that all the (classes of) inputs are taken into account. To define partial productivity indices, first some additional notation is necessary. Recall that all the items at the input side of our production unit are supposedly allocated to the following five, mutually disjunct, classes: capital (K ), labor (L), energy (E), materials (M), and services (S). The entire input price and quantity vectors can then be partitioned as w t = (w Kt , w Lt , w tE , w tM , wSt ) and x t = (x Kt , x Lt , x tE , x tM , x St ), respectively. Energy, materials, and services together form the category of intermediate inputs—that is, inputs which are acquired from other production units or imported. Capital and labor are called primary inputs. Consistent with this distinction, the price t t and quantity vectors can also be partitioned as w t = (w KL , w tEMS ) and x t = (x KL , x tEMS ), or t t t t t t t t as w = (w K , w L , w EMS ) and x = (x K , x L , x EMS ). Since monetary values are additive, total production cost can be decomposed in a number of ways, such as C t = ∑ wnt xnt + ∑ wnt xnt + ∑ wnt xnt + ∑ wnt xnt + ∑ wnt xnt n∈K

n∈L

t K

t L

t E

t K

t L

t EMS

n∈E

t M

≡C +C +C +C +C ≡C +C +C

t S

n∈M

n∈S

(2.58)

t t ≡ CKL + CEMS .

Based on the last line, total production cost change can be decomposed as

98 Bert M. Balk

0 0 C1EMS C1KL CEMS C1 CKL = + , (2.59) 0 0 C 0 C 0 CKL C 0 CEMS

but also as −1

−1 −1 C1EMS C1EMS C1 C1KL C1KL = + 1 0 , (2.60) 0 C 0 C1 CKL C C EMS

and then each ratio can be decomposed further. For instance, the labor-cost ratio can be decomposed as

CL1 = P (w1L , x1L , w L0 , x L0 )Q(w1L , x1L , w L0 , x L0 ) (2.61) CL0 ≡ PL (1, 0)QL (1, 0),,

for some pair of price and quantity indices. The labor productivity index for period 1 relative to period 0 is defined by

ILPROD(1, 0) ≡

QR (1, 0) ; (2.62) QL (1, 0)

that is, the ratio of an output quantity index to a labor input quantity index. Multiplying both sides of this definition by PR (1, 0) / PL (1, 0), using expressions (2.61), (2.30) and (2.29), and rearranging a bit delivers the following expression:

PL (1, 0) C1 / C1 ϒ 0 = ILPROD(1, 0) × L0 0 × 1 . (2.63) PR (1, 0) CL / C ϒ

It is interesting to look somewhat closer at the various components of this expression. First, the ratio PL (1, 0) / PR (1, 0) is usually called the real wage index. Its inverse is the counterpart of the total price recovery index, defined by expression (2.37). Second, CLt / C t (t = 0, 1) is the period t labor cost share, the share of labor cost in total input cost. Third, ϒ 0 / ϒ1 is the inverse profitability ratio. The practical usefulness of expression (2.63) lies in the conclusion that if profitability and the labor cost share remain constant, then real wage development must be proportional to labor productivity development. Notice that in expression (2.62) labor is not considered as a homogeneous commodity. The class L consists of several types of labor, each with their quantities xnt (say, hours worked) and prices wnt (say, wage per hour). Conventionally, however, labor is considered as homogeneous, so that the quantities xnt can be added up. Let Lt ≡ Σ n∈L xnt denote the total labor input of our production unit (say, measured in hours worked of

Empirical Productivity Indices and Indicators 99 whatever type) during period t (t = 0, 1). Then the simple labor productivity index for period 1 relative to period 0 is defined by QR (1, 0) . (2.64) L1 / L0 Formally, L1 / L0 is a simple sum or Dutot quantity index. The ratio of the two labor productivity indices,

ISLPROD(1, 0) ≡

LQUAL(1, 0) ≡

ISLPROD(1, 0) QL (1, 0) = 1 0 , (2.65) ILPROD(1, 0) L /L

is said to measure the shift in “labor quality”; actually this ratio measures the shift in the composition of labor input. To see this, let the labor input quantity index be defined as some weighted mean of the quantity relatives of the labor types; that is, QL (1, 0) ≡ Σ n∈L sn xn1 / xn0 with Σ n∈L sn = 1. Then it appears that x1 / L1 LQUAL(1, 0) = ∑ sn n0 0 . (2.66) xn / L n∈L t t Now xn / L is the share of the hours worked by type n ∈L in the total number of hours worked during period t, and LQUAL(1, 0) measures the aggregate shift of those shares from period 0 to period 1. Put otherwise, LQUAL(1, 0) measures the reallocation of labor input. The capital productivity index is defined as

IKPROD(1, 0) ≡

QR (1, 0) , (2.67) QK (1, 0)

where QK (1, 0) satisfies the equality C1K /CK0 = PK (1, 0)QK (1, 0). The other partial productivity indices IkPROD(1, 0) for k = E , M , S are defined similarly. The ratio of labor and capital productivity indices,

ILPROD(1, 0) QK (1, 0) = , (2.68) IKPROD(1, 0) QL (1, 0)

is called the index of capital deepening. Loosely speaking, this index measures the change of the quantity of capital input per unit of labor input. The relation between total factor (ITFPROD) and partial (IkPROD) productivity indices is not always simple. Let, for example, the total input quantity index QC (1, 0) be a two-stage index in which the second stage is a Fisher index,

QC (1, 0) ≡ Q F (Qk (1, 0), Ck1 , Ck0 ; k = K , L, E , M , S) (2.69)

100 Bert M. Balk and all the first-stage indices Qk (1, 0) are left unspecified. It is straightforward to check that then ITFPROD(1, 0) = =

QR (1, 0) QC (1, 0)

(Σ Q (1, 0)C k

k

QR (1, 0)

0 k

/C

Q (1, 0) Ck0 = Σk k 0 Q (1, 0) C

0

) (Σ Q (1, 0) 1/ 2

−1/ 2

R

k

k

−1

Ck1 / C1

QR (1, 0) Ck1 Σ k Q (1, 0) C1

)

−1/ 2

(2.70)

1/ 2

k

Σ C 0 (IkPROD(1, 0))−1 = k k C0

−1/ 2

Σ k Ck1 IkPROD(1, 0) C1

1/ 2

,

thus the TFP index is not a particularly simple function of the partial productivity indices. The first bracket shows a base-period-cost-share weighted harmonic mean, whereas the second bracket shows a comparison-period-cost-share weighted arithmetic mean. If instead as second-stage total input quantity index the Cobb-Douglas functional form, defined by expression (2.50), were chosen, then it appears that

ln ITFPROD(1, 0) = ∑ α k ln IkPROD(1, 0). (2.71) k

This is a very simple relation between TFP change and partial productivity change. Notice, however, that this simplicity comes at a cost. Definition (2.50) implies for the relation between aggregate and subaggregate input price indices that

PC (1, 0) = ∏ Pk (1, 0)αk k

C1 / C 0 . (2.72) Π(Ck1 / Ck0 )αk k

Such an index does not necessarily satisfy the fundamental Identity Test; that is, if all the prices in period 1 are the same as in period 0, then PC (1, 0) does not necessarily deliver as outcome 1. Though data requirements at the output side are the same, a partial productivity index needs less input data than a TFP index. For instance, for a labor productivity index, one only needs (more or less detailed) labor cost data. Partial productivity indices are useful for specific analytical purposes, but they don’t tell the whole story of productivity change. In an interesting assessment of the strengths and weaknesses of the various tools, Murray (2016, 124) concludes that “a complete understanding of productivity growth is best achieved by examining TFP and partial productivity measures together.”

Empirical Productivity Indices and Indicators 101

2.8. Indicators Let us now turn to profit and its development through time.13 This is naturally measured by the difference Π1 − Π 0 . Recall that such a difference only makes sense when the two money amounts involved, profit from period 0 and profit from period 1, are deflated by some general inflation measure (such as the headline CPI), or when the two periods are so close that deflation is deemed unnecessary. In the remainder of this chapter, when discussing difference measures, either of the two situations is tacitly presupposed. How does one decompose the profit difference into a price and a quantity component? By noticing that

Π1 − Π 0 = (R1 − R0 ) − (C1 − C 0 ), (2.73)

we see that the question reduces to the question of how to decompose revenue change R1 − R0 and cost change C1 − C 0 into two parts. This is where indicators come into play. Let us again take revenue as example. Formally, a bilateral price indicator is a continuously differentiable function ( p1 , y1 , p0 , y 0 ) : ℜ4+M+ → ℜ that, except in extreme situations, correctly indicates any increase or decrease of the elements of the price vectors p1 or p0, conditional on the quantity vectors y1 and y 0. Likewise, a bilateral quantity indicator is a continuously differentiable function of the same variables, ( p1 , y1 , p0 , y 0 ) : ℜ4+M+ → ℜ, that, except in extreme situations, correctly indicates any increase or decrease of the elements of the quantity vectors y1 or y 0, conditional on the price vectors p1 and p0. The number M is called the dimension of the price or quantity indicator. Notice that both functions may take on negative or zero values. The basic requirements on price and quantity indicators are analogous to those for indices and comprise the following: 1. that they exhibit the desired monotonicity properties globally; 2. that they satisfy the (analogue of the) Identity Test; that is, ( p0 , y1 , p0 , y 0 ) = 0 and ( p1 , y 0 , p0 , y 0 ) = 0 ; 3. that they are homogeneous of degree 1 in prices (quantities, respectively); 4. that they are invariant to changes in the units of measurement of the items. The reader is referred to Balk (2008, 126–129) for the formalization and an extensive discussion of these and other requirements. Though ( p1 , y1 , p0 , y 0 ) and( p1 , y1 , p0 , y 0 ) are defined as functions of 4M variables, it appears that any such function that is invariant to changes in the units of measurement can be written as a function of only 3M variables, namely the price relatives

102 Bert M. Balk pm1 / pm0 or the quantity relatives y1m / ym0 , the comparison period values v1m ≡ pm1 y1m, and the base period values vm0 ≡ pm0 ym0 (m = 1, , M ). The interesting point to notice here is that differences pm1 − pm0 or y1m − ym0 are not necessary for the computation of ( p1 , y1 , p0 , y 0 ) or ( p1 , y1 , p0 , y 0 ). Also, here some simple examples might be more useful than formal definitions. The Laspeyres price indicator as function of prices and quantities is defined as L ( p1 , y1 , p0 , y 0 ) ≡ ( p1 − p0 ) ⋅ y 0 .

This indicator can be written as a function of individual price relatives and (base period) values, but also as a function of the Laspeyres price index and aggregate base period value,

M

(

)

L ( p1 , y1 , p0 , y 0 ) = ∑ ( pm1 / pm0 − 1)vm0 = P L ( p1 , y1 , p0 , y 0 ) − 1 V 0 , m =1

where we recall that V 0 ≡ Σ mM=1vm0 . Similarly, the Paasche price indicator P ( p1 , y1 , p0 , y 0 ) ≡ ( p1 − p0 ) ⋅ y1

can be written as a function of individual price relatives and (comparison period) values, and as a function of the Paasche price index and aggregate comparison period value,

M

(

)

P ( p1 , y1 , p0 , y 0 ) = ∑ (1 − pm0 / pm1 )v1m = 1 − 1 / P P ( p1 , y1 , p0 , y 0 ) V 1 , m =1

where V 1 ≡ Σ mM=1v1m. Notice the reciprocals here. Finally, the Bennet price indicator is defined as

B ( p1 , y1 , p0 , y 0 ) ≡ (1 / 2)( p1 − p0 ) ⋅ ( y 0 + y1 ),

which is the arithmetic mean of L ( p1 , y1 , p0 , y 0 ) and P ( p1 , y1 , p0 , y 0 ). One immediately verifies that M M B ( p1 , y1 , p0 , y 0 ) = (1 / 2) ∑ ( pm1 / pm0 − 1)vm0 + ∑ (1 − pm0 / pm1 )v1m m =1 m =1 L P 1 1 1 0 0 0 1 1 0 0 = (1 / 2) P ( p , y , p , y ) − 1 V + 1 − 1 / P ( p , y , p , y ) V .

(

)

(

)

Notice that the Bennet indicator is symmetric in the two periods, and hence satisfies the (analogue of the) Time Reversal Test. The Bennet indicator is the analogue of the Marshall-Edgeworth index as well as the Fisher index.

Empirical Productivity Indices and Indicators 103 The analogue of the Product Test requires that price indicator plus quantity indicator equals the value difference. Put otherwise, the analogue of deflation is subtraction, and one easily checks that the following relations hold:

V 1 − V 0 − L ( p1 , y1 , p0 , y 0 ) = P ( p1 , y1 , p0 , y 0 )

V 1 − V 0 − P ( p1 , y1 , p0 , y 0 ) = L ( p1 , y1 , p0 , y 0 )

V 1 − V 0 − B ( p1 , y1 , p0 , y 0 ) = B ( p1 , y1 , p0 , y 0 ),

where P (.), L (.), and B (.) are the Paasche, Laspeyres, and Bennet quantity indicators, respectively. The last equation expresses that the Bennet indicators are ideal; that is, they satisfy the analogue of the Product Test and have the same functional form. Finally, a noteworthy feature of price and quantity indicators is that, due to their additive structure, aggregation is not an issue. Put otherwise, indicators are always consistent in aggregation.

2.9. Decomposing a Profit Difference Continuing now from expression (2.73), we select pairs of price and quantity indicators such that

R1 − R0 = ( p1 , y1 , p0 , y 0 ) + ( p1 , y1 , p0 , y 0 ) (2.74) ≡ R (1, 0) + R (1, 0),

and similarly,

C1 − C 0 = (w1 , x1 , w 0 , x 0 ) + (w1 , x1 , w 0 , x 0 ) ≡ C (1, 0) + C (1, 0).

(2.75)

Notice that the dimension of the indicators in these two decompositions is generally different. Also the functional form of the indicators may or may not be the same. The profit difference can then be written as Π1 − Π 0

= R (1, 0) + R (1, 0) − [C (1, 0) + C (1, 0)] (2.76) = R (1, 0) − C (1, 0) + R (1, 0) −C (1, 0).

The first two terms at the right-hand side of the last equality sign provide the price component, whereas the last two terms provide the quantity component

104 Bert M. Balk of the profit difference. Thus, based on this decomposition, the TFP indicator is defined by

DTFPROD(1, 0) ≡ R (1, 0) − C (1, 0); (2.77)

that is, output quantity indicator minus input quantity indicator. Notice that productivity change is now measured as an amount of money. An amount greater (less) than 0 indicates productivity increase (decrease).14 The equivalent expressions for difference-type TFP change are

DTFPROD(1, 0) = [R1 − R0 − R (1, 0)] − [C1 − C 0 − C (1, 0)] (2.78)

= [R1 − R0 − R (1, 0)] − C (1, 0) (2.79)

= R (1, 0) − [C1 − C 0 − C (1, 0)], (2.80)

which can be useful in different situations. The total price recovery (TPR) indicator is defined as

DTPR(1, 0) ≡ R (1, 0) − C (1, 0). (2.81)

This is also an amount of money. It is clear that Π1 − Π 0 = DTFPROD(1, 0) + DTPR(1, 0), and that, if Π1 = Π 0 then

DTFPROD(1, 0) = − DTPR(1, 0). (2.82)

For a non-market production unit, a productivity indicator is difficult to define. Though one might be able to construe some output quantity indicator, it is hard to see how, in the absence of output prices, such an indicator could be given a money dimension. Can we also define partial productivity indicators? Given a set of indicators, the laborcost difference between periods 0 and 1 is decomposed as

C1L − CL0 = (w1L , x1L , w L0 , x L0 ) + (w1L , x1L , w L0 , x L0 ) (2.83) ≡ L (1, 0) + L (1, 0).

In the same way one can decompose the capital, energy, materials, and services cost difference. However, since costs are additive, it turns out that the TFP indicator can be written as

DTFPROD(1, 0) = R (1, 0) −

∑

k = K ,L , E , M ,S

k (1, 0). (2.84)

By definition, the left-hand side measures real profit change. The right-hand side provides the contributing factors. The contribution of category k to real profit change is simply measured by the amount k (1, 0). A positive amount, which means that the aggregate quantity of input category k has increased, means a negative contribution to real profit change.

Empirical Productivity Indices and Indicators 105 Ball et al. (2015) called R (1, 0) − E (1, 0) a partial energy productivity indicator. However, notice that summing all those partial indicators, Σ k = K , L , E , M ,S (R (1, 0) − k (1, 0)), does not lead to the overall indicator, DTFPROD(1, 0), which makes the interpretation of such partial indicators difficult. In the extended online version of this chapter, the equivalence of the multiplicative and additive models is discussed. First, it is shown that, given that the profitability ratio can be decomposed into two factors, ϒ1 / ϒ 0 = ITFPROD(1, 0) ×ITPR(1, 0), the profit difference Π1 − Π 0 decomposes into four factors, namely a productivity effect, a price recovery effect, and two margin effects. The existence of these margin effects “makes the difference-based model of profit change richer than the ratio-based model of profitability change,” according to Grifell-Tatjé and Lovell (2015, 201). However, it is also shown that, given that the profit difference can be decomposed into two factors, Π1 − Π 0 = DTFPROD(1, 0) + DTPR(1, 0), the profitability ratio ϒ1 / ϒ 0 decomposes into four factors, namely a productivity effect, a price recovery effect, and two margin effects. The overall conclusion, then, is that the two models, the multiplicative model based on profitability and the additive model based on profit, are completely equivalent.

2.10. The KL-VA Model (1) There are alternatives to the KLEMS-Y input–output model. The most prominent of these models, widely employed by economists, uses value added (VA) as its output concept. The production unit’s value added (VA) is defined as its revenue minus the costs of energy, materials, and services; that is,

t VAt ≡ Rt − CEMS (2.85) = pt ⋅ y t − w tEMS ⋅ x tEMS .

The value-added concept subtracts the total cost of intermediate inputs from the revenue obtained, and in doing so essentially conceives the production unit as producing value added (that is, money) from the two primary input categories, capital and labor. It is assumed that VAt > 0.15 The fundamental accounting relation of our production unit was given by expression (2.3) as C t + Πt = Rt . Subtracting from the left-and the right-hand side intermediate t inputs cost CEMS , using expressions (2.58) and (2.85), this accounting relation transforms t t into CKL + Π = VAt . Thus the same production unit can be described by two different accounting relations, featuring different input and output concepts. Although gross output, represented by the quantity vector y t, is the natural output concept, the value-added concept is important when one wants to aggregate single units to larger entities. Gross output consists of deliveries to final demand and intermediate

106 Bert M. Balk destinations. The split between these two output categories depends very much on the level of aggregation. Value added is immune to this problem. It enables one to compare (units belonging to) different industries. From a welfare-theoretic point of view, the value-added concept is important because value added can be conceived as the income (from production) that flows into society. In the KL-VA input–output model, it is natural to define (value-added-based) profitability as the ratio of value added to primary inputs cost,

t t ΓVA ≡ VAt / CKL , (2.86)

and the natural starting point for defining a productivity index is to consider the development of this ratio through time. Since

Γ1VA VA1 / VA0 = 1 , (2.87) 0 0 ΓVA CKL / CKL

we need a decomposition of the value-added ratio and a decomposition of the primary inputs cost ratio. The question of how to decompose a value-added ratio into a price and a quantity component cannot be answered unequivocally. There are several options here, the technical details of which are discussed in the extended online version of this chapter. It is there shown that • single deflation is simple, leads to a quantity index that is always positive, but this index may be biased; • double deflation by a Fisher index leads to a quantity index that is unbiased but may not always be well defined; • double deflation by a Montgomery-Vartia index leads to a quantity index that is always well defined but fails the Equality Test. Thus, there is no theoretically entirely satisfactory solution to the problem of decomposing a value-added ratio into a price index and a quantity index.

2.11. The KL-VA Model (2) Suppose now that a practically satisfactory decomposition of the value-added ratio is available; that is,

VA1 = PVA (1, 0)QVA (1, 0), (2.88) VA0

Empirical Productivity Indices and Indicators 107 and that, by employing some pair of price and quantity indices, the primary inputs cost ratio is decomposed as

C1KL 0 0 0 0 = P (w1KL , x1KL , w KL , x KL ) Q(w1KL , x1KL , w KL , x KL ) 0 CKL (2.89) ≡ PKL (11, 0) QKL (1, 0).

The value-added-based TFP index for period 1 relative to period 0 is then defined as

ITFPRODVA (1, 0) ≡

QVA (1, 0) . (2.90) QKL (1, 0)

This index measures the “quantity” change of value added relative to the quantity change of primary inputs, or, can be seen as the index of real value added relative to the index of real primary inputs. As observed in section 2.11, profit in the KL-VA model is the same as profit in the KLEMS-Y model, and this also applies to profit differences and their price and quantity components. One easily checks that DTFPRODVA (1, 0) ≡ VA (1, 0) − KL (1, 0)

= R (1, 0) − C (1, 0) (2.91) = DTFPROD(1, 0);

that is, the productivity indicators are the same in the two models. This, however, does not hold for the productivity indices. One usually finds that ITFPRODVA (1, 0) ≠ ITFPROD(1, 0). Balk (2009) showed that if profit is zero in both periods, that is, Rt = C t (t = 0, 1), then, for certain two-stage indices that are second- order differential approximations to Fisher indices,

ln ITFPRODVA (1, 0) = D(1, 0)ln ITFPROD(1, 0), (2.92)

where D(1, 0) ≥ 1 is the (mean) Domar-factor (= ratio of revenue over value added). Usually expression (2.92) is, in a continuous-time setting, derived under a set of strong neoclassical assumptions (see, for instance, Gollop 1979; Jorgenson et al. 2005, 298; Schreyer 2001, 143), so that this equation seems to be some deep economic-theoretical result. From the foregoing it may be concluded, however, that the inequality of the value-added-based productivity index and the gross-output-based productivity index is only due to the mathematics of ratios and differences. There is no underlying economic phenomenon, which does not mean that there cannot be given an economic interpretation to equation (2.92). The value-added-based labor productivity index for period 1 relative to period 0 is defined as

ILPRODVA (1, 0) ≡

QVA (1, 0) , (2.93) QL (1, 0)

108 Bert M. Balk where QL (1, 0) was defined by expression (2.61). The index defined by expression (2.93) measures the “quantity” change of value added relative to the quantity change of labor input; or, this expression can be seen as the index of real value added relative to the index of real labor input. Recall that the labor quantity index QL (1, 0) is here defined as acting on the prices and quantities of all the types of labor that are being distinguished. Suppose that the units of measurement of the various types are in some sense the same; that is, the quantities of all the types are measured in hours, or in full-time equivalent jobs, or in some other common unit. The simple value-added-based labor productivity index, defined as QVA (1, 0) , (2.94) L1 / L0 can then be interpreted as an index of real value added per unit of labor. Recall that Lt ≡ Σ n∈L xnt (t = 0, 1). The simple value-added-based labor productivity index frequently figures at the left-hand side (thus, as explanandum) in a growth accounting equation. However, for deriving such a relation, nothing spectacular is needed, as will now be shown. Consider the definition of the value-added-based TFP index, expression (2.90), and rewrite this as

ISLPRODVA (1, 0) ≡

QVA (1, 0) = ITFPRODVA (1, 0) × QKL (1, 0). (2.95)

Dividing both sides of this equation by the Dutot labor quantity index L1 / L0, applying definitions (2.65) and (2.94), one obtains

ISLPRODVA (1, 0) = ITFPRODVA (1, 0) ×

QKL (1, 0) × LQUAL(1, 0). (2.96) QL (1, 0)

Taking logarithms and, on the assumption that all the index numbers are in the neighborhood of 1, interpreting these as percentages, the last equation can be read as the following: (simple) labor productivity growth equals TFP growth plus “capital deepening” plus “labor quality” growth.16 Again, productivity change is measured as a residual and, thus, the three factors at the right-hand side of the previous equation can in no way be regarded as causal factors. If, continuing the example of section 2.6.2, the primary inputs quantity index were defined as a two-stage index of the form

QKL (1, 0) ≡ QK (1, 0)α QL (1, 0)1− α (0 < α < 1), (2.97)

where any reader will recognize the simple Cobb-Douglas form, then the index of “capital deepening” reduces to the particularly simple form α

QKL (1, 0) QK (1, 0) = . (2.98) QL (1, 0) QL (1, 0)

Empirical Productivity Indices and Indicators 109 As mentioned earlier, the “labor quality” index, LQUAL(1, 0), basically measures compositional shift or structural change among the various labor input types, since it is a ratio of two quantity indices.

2.12. Aggregation In the previous sections we considered the measurement of productivity change for a single, consolidated production unit. This section continues by considering an ensemble of such units, and the relation between aggregate and subaggregate (or individual) measures of productivity change. The theory surveyed here is applicable to a variety of situations, such as (1) a firm consisting of a number of plants, (2) an industry consisting of a number of firms, or (3) an economy or, more precisely, the commercial sector of an economy consisting of a number of industries. On an intuitive level, the relation between aggregate and subaggregate, or individual, measures of productivity change is not too difficult to understand. Productivity is output quantity divided by input quantity and, thus, productivity change is output quantity change divided by input quantity change. Any aggregate is somehow the sum of its parts, which in the present context implies that aggregate productivity is somehow a weighted mean of subaggregate (or individual) productivities, where the weights somehow express the “importance” of the subaggregates, or individual units, making up the aggregate. Hence, there are two, independent, factors responsible for aggregate productivity change: (1) productivity change at subaggregate, or individual, level; and (2) change of the “importance” of the subaggregate, or individual, units. Coming down to practice, things are rapidly becoming complicated. First, firms produce and use multiple commodities, which brings input and output prices into play. Quantities of different commodities cannot be added, but must be aggregated by means of prices. Through time, prices are also changing, which implies that we cannot simply talk about aggregate prices and quantities, but must talk about price and quantity index numbers. Moreover, aggregation rules are not unique anymore. Second, firms and industries deliver to each other, which implies that the “simple” addition of production units easily leads to a form of double-counting of outputs and inputs. Third, especially when we are dealing with firms or plants, an important fact to take into account is the dynamics of growth, decline, birth, and death of production units. Let denote an ensemble of consolidated production units. For each unit k and time period t the KL-VA accounting identity in nominal values reads

kt CKL + Π kt = VAkt (k ∈), (2.99)

where we recall that by assumption VAkt > 0 for all production units and time periods considered.

110 Bert M. Balk The KL-VA model is chosen because capital input, labor input, and value added are specific for each production unit. Hence, adding the relations (2.99) over all the units,

∑C

k ∈

kt KL

+ ∑ Π kt = ∑ VAkt , (2.100) k ∈

k ∈

avoids the possibility of double-counting and delivers the same relation for the ensemble considered as a consolidated production unit: t CKL + Π t = VAt , (2.101)

t kt with CKL ≡ Σ k ∈CKL , Π t ≡ Σ k ∈ Π kt , and VAt = Σ k ∈VAkt . The ensemble we consider is dynamic; that is, its composition changes through time. Thus, wherever necessary, we must add a superscript t to . For any two time periods compared, whether adjacent or not, a distinction must then be made between continuing, exiting, and entering production units. In particular,

0 = 01 ∪ 0 (2.102)

1 = 01 ∪ 1 , (2.103)

where 01 denotes the subset of continuing units (active in both periods), 0 the subset of exiting units (active in the base period only), and 1 the subset of entering units (active in the comparison period only).17 The sets 01 and 0 are disjunct, as are 01 and 1. It is important to observe that in any application the distinction between continuing, entering, and exiting production units depends on the length of the time periods being compared, and on the time span between these periods. Of course, when the production units studied form a balanced panel, then the sets 0 and 1 are empty. The same holds for the case where the production units are industries.

2.12.1. Top-Down We first consider TFP indices. For the static case— 0 and 1 empty, so that 0 = 1 = — Balk (2015) studied the relation between the aggregate TFP index ITFPRODVA (1, 0) k and the subaggregate (individual or industrial) TFP indices ITFPRODVA (1, 0) (k ∈), all the indices defined according to expression (2.90). Out of a number of alternatives, he recommended18 k ln ITFPRODVA (1, 0) = ∑ ψ k ln ITFPRODVA (1, 0) k ∈

k P k (1, 0) PKL (1, 0) ln − + ∑ ψ k ln VA (2.104) PKL (1, 0) PVA (1, 0) k ∈ C k1 / C k 0 + ∑ ψ k ln KL1 KL , 0 CKL / CKL k ∈

Empirical Productivity Indices and Indicators 111 where

VAk1 VAk 0 LM 1 , 0 VA VA ψk ≡ VAk1 VAk 0 Σ LM 1 , 0 k ∈ VA VA

(k ∈)

is the (logarithmic) mean share of production unit k’s value added in aggregate value added. There occur three main terms at the right-hand side of expression (2.104). The first is a weighted mean of unit-specific TFP changes. The second is the aggregate effect of differential price change at the output and input side of the production units. The third can be interpreted as the aggregate effect of relative size change, where the size of a unit is measured by its primary-input cost share. If there is no differential price change, then the second term at the right-hand side vanishes.19 Moreover, if for all units k ∈ and time periods t = 0, 1 value added equals kt kt t primary input cost, VAkt = CKL , or profit Π kt = 0, then CKL / CKL = VAkt / VAt (t = 0, 1), and by using the definition of the logarithmic mean, one easily checks that also the third term vanishes. Under these two conditions, expression (2.104) reduces to

k ln ITFPRODVA (1, 0) = ∑ ψ k ln ITFPRODVA (1, 0). (2.105) k ∈

0

1

For the dynamic case— and non-empty—the decomposition in expression (2.104) must be extended. For the details the reader is referred to Balk (2015, ex. [79]). For aggregate and subaggregate simple labor productivity indices, Balk (2014) and Dumagan and Balk (2016) derived the following relation for the static case:20 k ln ISLPRODVA (1, 0) = ∑ ψ k ln ISLPRODVA (1, 0)

k ∈

P k (1, 0) Lk1 / L1 (2.106) k + ∑ ψ k ln VA ψ ln + ∑ Lk 0 / L 0 , PVA (1, 0) k ∈ k ∈

k where ISLPRODVA (1, 0) and ISLPRODVA (1, 0) (k ∈) were defined according to expression (2.94). The first term at the right-hand side of the equality sign is a weighted mean of unit-specific simple labor productivity changes. The second term is a weighted mean k of individual relative output price changes PVA (1, 0) / PVA (1, 0) ((k ∈)). The third term measures labor reallocation.

2.12.2. Bottom-Up The bottom-up approach, as surveyed by Balk (2016), has two main characteristics. The first is that one freely talks about levels, of input, output, or productivity. The second is that one does not care too much about the interpretation of aggregate input, output, or productivity levels.

112 Bert M. Balk Let us first turn to the concept of level. For each production unit k ∈t real value added is (ideally) defined as

k RVAk (t , b) ≡ VAkt / PVA (t , b); (2.107)

that is, nominal value added at period t divided by (or, as one says, deflated by) a production-unit- k-specific value-added based price index for period t relative to a certain reference period b, where period b may or may not precede period 0. Notice that this definition tacitly assumes that production unit k, existing in period t, also existed or still exists in period b; otherwise, deflation by a production-unit-k-specific index would be impossible. When production unit k does not exist in period b, then for deflation some substitute index must be used. It is good to realize that, by using the Product Test for value-added-based price k and quantity indices, RVAk (t , b) = VAkbQVA (t , b). Put otherwise, real value added is a (normalized) quantity index. Like real value added, real primary, or capital-and-labor, input, relative to reference period b, is (ideally) defined as deflated primary input cost,

k kt k X KL (t , b) ≡ CKL / PKL (t , b), (2.108)

and real labor input, relative to reference period b, is (ideally) defined as deflated labor cost,

X Lk (t , b) ≡ CLkt / PLk (t , b). (2.109)

Using the foregoing building blocks, the value-added-based TFP level of production unit k at period t is defined as real value added divided by real primary input,

k TFPRODVA (t , b) ≡

RVAk (t , b) . (2.110) k X KL (t , b)

Notice that numerator and denominator are expressed in the same price level, namely k that of period b. Thus TFPRODVA (t , b) is a dimensionless variable. Likewise, the value-added-based labor productivity level of unit k at period t is defined as real value added divided by real labor input,

k LPRODVA (t , b) ≡

RVAk (t , b) . (2.111) X Lk (t , b)

This is also a dimensionless variable. Finally, the simple value-added-based labor productivity level of unit k at period t is defined by RVAk (t , b) . (2.112) Lkt Notice that this level does have a dimension, namely money-of-period-b per unit of labor (say, hour of whatever type of work).

k SLPRODVA (t , b) ≡

Empirical Productivity Indices and Indicators 113 Now, let PROD kt denote (the logarithm or inverse of) total factor productivity k k TFPRODVA (t , b), labor productivity LPRODVA (t , b), or simple labor productivity k SLPRODVA (t , b), then the bottom-up approach is concerned with the decomposition of PROD1 − PROD 0 ≡

∑θ

k1

PROD k1 −

1

k ∈

∑θ

k ∈

k0

PROD k 0 , (2.113)

0

where the period-specific weights, measuring the relative size (importance) of the prok0 k1 duction units, add up to 1; that is, Σ k ∈0 θ = Σ k ∈1 θ = 1. Balk’s (2016) survey revealed the existence of a large number of decompositions, scattered through the literature. Because of its symmetry and its natural benchmarks for exiting and entering production units, the Diewert-Fox (2010) decomposition was preferred. This decomposition reads PROD1 − PROD 0

(

1 01 = ∑ θk1 PROD − PROD 1 k ∈ 1

)

+

θ k 0 + θ k1 PROD k1 − PROD k 0 2 k ∈ 01

+

∑ (θ

(

∑

k ∈ 01

k1

)

(2.114)

PROD k 0 + PROD k1 − θ k 0 ) − a 2

(

)

0 01 − ∑ θ k 0 PROD − PROD 0 , k ∈ 0 where PROD ≡ Σ k ∈ 0 θk 0 PROD k 0 / Σ k ∈ 0 θk 0 is the mean productivity level of the 0

exiting units; PROD ≡ Σ k ∈ 1 θk1 PROD k1 / Σ k ∈ 1 θk1 is the mean productivity level of the 1

entering units; and PROD t ≡ Σ k ∈ 01 θkt PROD kt / Σ k ∈ 01 θkt is the aggregate productivity level of the continuing production units at period t (t = 0, 1); the relative size of continuing units is defined by θ kt ≡ θkt / Σ k ∈ 01 θkt (k ∈ 01 ; t = 0, 1); and a is an arbitrary scalar. The first right-hand side term of expression (2.114) refers to the entering production units. As we see, its magnitude is determined by the period 1 share of the entrants and the productivity gap with the continuing units. The last right- hand side term refers to the exiting production units. The magnitude of this term depends on the share of the exiters and the productivity gap with the continuing units. The second and third term refer to the continuing production units. These units may contribute positively in two ways: if their productivity levels on average increase, or if the units with mean productivity levels above (or below) the scalar a increase (or decrease) in relative size. Notice that the third term is the only place where an arbitrary scalar a can be inserted, since the relative weights of the continuing production units add up to 1 in both periods. 01

114 Bert M. Balk Though the term itself is invariant to the actual magnitude of a, the unit-specific components (θk1 − θ k 0 )((PROD k 0 + PROD k1 ) / 2 − a) are not. The Diewert-Fox decomposition (2.114) bears a strong resemblance to the decomposition proposed decades ago by Griliches and Regev (1995), which has found many applications. The last decomposition reads: PROD1 − PROD 0

(

1 = ∑ θk1 PROD − a k ∈ 1

)

θ k 0 + θ k1 PROD k1 − PROD k 0 2 01 k ∈

(

+∑

)

(2.115)

PROD k 0 + PROD k1 + ∑ (θk1 − θk 0 ) − a 2 k ∈ 01

(

)

0 − ∑ θk 0 PROD − a . k ∈ 0 With respect to the scalar a there are several options available in the literature. A rather natural choice is a = (PROD 0 + PROD1 ) / 2, the overall two-period mean aggregate productivity level. Then, entering units contribute positively to aggregate productivity change if their mean productivity level is above this overall mean. Exiting units contribute positively if their mean productivity level is below the overall mean. Continuing units can contribute positively in two ways: if their productivity level increases, or if the units with productivity levels above (below) the overall mean increase (decrease) in relative size. As we see, the single benchmark in the Griliches-Regev decomposition is replaced by a threefold benchmark in the Diewert-Fox decomposition. If there are no exiting or entering units, that is, 0 = 1 = 01, then both decompositions reduce to PROD1 − PROD 0 =

θ k 0 + θ k1 PROD k1 − PROD k 0 2 k ∈ 01

(

∑

)

PROD k 0 + PROD k1 − a , + ∑ (θ − θ ) 2 k ∈ 01 k1

k0

where we recognize a Bennet-type decomposition. The question of which weights θkt are appropriate, when a choice has been made as to the productivity levels PROD kt (k ∈t ), has received some attention in the literature. Given that somehow PROD kt is output divided by input, should θkt be output-or input-based? And how is this related to the type of mean—arithmetic, geometric, or harmonic? The literature does not provide us with definitive answers. Indeed, as long

Empirical Productivity Indices and Indicators 115 as one stays in the bottom-up framework, it is unlikely that a convincing answer can be obtained. This is the point where the bottom-up view must connect with the top-down view. For details the reader is referred to Balk (2018).

2.13. Concluding Remarks The World Confederation of Productivity Science at its website (www.wcps.info) once defined productivity as “the single ratio of output to input—improving productivity means getting the most out for the least put in. It is important to companies, to nations (since it is a major determinant of national competitiveness) and to the world.” Basically this chapter reviews the empirical implementation of this “single ratio” concept. It has been shown that there are a number of alternative, non-competing measurement objectives; that one can distinguish between total or partial measures; and that there is choice with respect to functional forms. It is useful to recapitulate the main characteristics of the approach outlined in the foregoing: • We did not make any of the usual neoclassical structural and behavioral assumptions; in particular, we did not assume the existence of production functions or a certain kind of optimizing behavior of the production units. Nevertheless, the approach advocated here is able to deliver the same results as the neoclassical approach; for example, it appears that for a table in which output growth is decomposed into the contributions of TFP change and input growth, the neoclassical assumptions are neither necessary nor do they contribute anything to our understanding of what productivity precisely is. • To get insight into the “drivers” of productivity change at the lowest level of aggregation, appropriate theoretical instruments as well as additional data must be invoked. This, however, brings us outside the area of empirical measurement of productivity change as such. • The data required here are in the first place nominal value data extracted from the annual ex post profit/loss accounts of a production unit. The extent to which the various cost categories are covered (capital, labor, energy, materials, services) determines whether one can compile total factor, multi-factor, or partial productivity indices. • For the various monetary flows considered, at the input and output sides of a production unit, we need bespoke deflators. As shown, for example, deflating nominal value added by a revenue-based price index is likely to generate a picture of output growth contaminated by price effects. Similarly, deflating input or output nominal value changes of individual production units by industry-level price indices is likely to generate bias.

116 Bert M. Balk • Though nominal values can be observed “as a whole,” the price indices serving as deflators are functions of all the detailed price and quantity data constituting the monetary flows. In practice, such price indices are compiled from samples of relative price and value data. Put otherwise, any deflator tacitly comes with a certain (sampling) inaccuracy—a fact hardly acknowledged in discussions of particular numerical outcomes. • It should also be noted that there is a certain freedom in the choice of functional form for the deflators. One should be aware, however, that any choice at the “price side” has implications for the “quantity side.” • As a matter of convenience, the theory reviewed here is cast in terms of time: a production unit or an ensemble of such units is moving from one period to the next. With a little bit of imagination, the theory can be modified to be applicable to cross- sectional comparisons. Of course, there are some limitations here: a plant or firm cannot be at different locations at the same time. For industries, however, international comparisons are feasible, provided that the underlying data are comparable. • The theory reviewed here assumes that at the (physical) output side, market prices are available. It is possible, however, to extend the theory by including unpriced bad outputs with negative shadow prices.

Notes 1. An extended version of this chapter is available at SSRN: http://ssrn.com/abstract=2776956. This chapter is not an intellectual history of the subject. The references are not intended as originality claims but serve to direct the reader to quickly available sources for further reading or for details that had to be omitted due to space limitations. 2. The precise definitions of these classes are beyond the scope of this chapter. 3. Other input–output models are discussed in Balk (2010). For an empirical comparison the reader is referred to Vancauteren et al. (2012). 4. It is not so simple to link this definition to the various profit concepts discussed by Marr (2012). The closest comes his concept of “economic value added,” since the (user) cost of (tangible and intangible) capital assets is included in C t . Marr’s “pretax profit” (EBT or EBITDA) t corresponds to revenue minus the cost of labor, energy, materials, and services, Rt − CLEMS ,a t concept Balk (2010) called cash flow, CF . Marr’s “net profit” then equals cash flow minus tax on production. Our definition of profit includes cost and revenue of raising and maintaining financial capital. See Diewert (2014) for an attempt to disentangle these components. 5. For any two strictly positive real numbers a and b, their logarithmic mean is defined by LM (a, b) = (a − b) / ln(a / b) if a ≠ b and LM (a, a) = a. It has the following properties: (1) min(a, b) ≤ LM (a, b) ≤ max(a, b); (2) LM (a, b) is conLM (λa, λb) = λLM (a, b) (λ > 0); LM (a, b) = LM (b, a); tinuous; (3) (4) (5) (ab)1/2 ≤ LM (a, b) ≤ (a + b) / 2 ; (6) LM (a, 1) is concave. See Balk (2008, 134–136) for details. 6. Marr (2012) distinguishes here between three concepts: “net profit margin” (based on cash t flow minus tax on production); “gross profit margin” (where CLEMS covers only the direct costs of production and distribution of goods and services); and “operating profit margin” t (where Rt and CLEMS cover only regular operations of the production unit).

Empirical Productivity Indices and Indicators 117 7. The revenue growth rate is the 6th of the 75 key performance indicators discussed by Marr (2012). 8. Under extreme conditions the monotonicity properties of geometric mean indices break down; see Balk (2008, 72). 9. This approach follows Diewert (1992), Diewert and Nakamura (2003), and Balk (2003). Notice that the productivity concept introduced here is purely descriptive. As components of productivity change may be distinguished between technological change, efficiency change, scale effects, and input/output-mix effects; economic research has unearthed a large number of underlying drivers. Chapter 4 of this Handbook discusses a narrower concept of productivity. 10. But note that it is not necessary to assume that Rt = C t (t = 0,1) , as Dumagan and Ball did. 11. Gordon (2016), for instance, employs a Cobb-Douglas function with α K = 0.3 and α L = 0.7. 12. This point was also forcefully made by Felipe and McCombie (2013). 13. The extended online version of this chapter contains a section on the profit-revenue and profit-cost margin ratios. 14. This approach follows Balk (2003). 15. An early advocate of the value-added output concept was Burns (1930). Specifically, he favored net value added (see Balk 2010 for the definition). Burns was aware of the possibility that for very narrowly defined production units and small time periods, value added may become non-positive. 16. This decomposition repeatedly turns up in the (labor productivity) growth accounting literature, a recent example being Muntean (2014). This author used an endogenous rate of return for capital, so that profit Πt = 0 . Though the neoclassical growth accounting theory was invoked, this theory neither proved necessary—since all the empirical results remained within the confines of the approach sketched here—nor helpful in understanding the “sources” of TFP change. 17. The definition of “active” should of course be made precise in any empirical application on microdata to ascertain that exit of a certain production unit is due to quitting business and not due to, for example, obtaining a new name, merging with another unit, splitting into two or more units, falling below the observation threshold; likewise for entry. 18. See Diewert (2016) for the asymmetric, base period weighted, variant which is known as the Generalized Exactly Additive Decomposition (GEAD). Calver and Murray (2016) provide an interesting application on Canadian data. 19. If there is differential price change at the output side but not at the input side, then the GEAD reduces to the so-called CSLS decomposition. See Calver and Murray (2016) for an example of the dramatic differences occurring between these two decomposition methods when there does exist differential price change. 20. See Diewert (2016) for the asymmetric, base period weighted variant, also known as the GEAD.

References Balk, B. M. 2003. “The Residual: On Monitoring and Benchmarking Firms, Industries, and Economies with Respect to Productivity.” Journal of Productivity Analysis 20: 5–47. Reprinted in National Accounting and Economic Growth, edited by John M. Hartwick. The International Library of Critical Writings in Economics No. 313. Cheltenham, UK; Northampton, MA: Edward Elgar, 2016.

118 Bert M. Balk Balk, B. M. 2008. Price and Quantity Index Numbers: Models for Measuring Aggregate Change and Difference. New York: Cambridge University Press. Balk, B. M. 2009. “On the Relationship Between Gross-Output and Value-Added Based Productivity Measures: The Importance of the Domar Factor.” Macroeconomic Dynamics 13(S2): 241–267. Balk, B. M. 2010. “An Assumption-Free Framework for Measuring Productivity Change.” The Review of Income and Wealth 56(Special Issue 1): S224–S256. Reprinted in National Accounting and Economic Growth, edited by John M. Hartwick. The International Library of Critical Writings in Economics No. 313. Cheltenham, UK; Northampton, MA: Edward Elgar, 2016. Balk, B. M. 2014. “Dissecting Aggregate Output and Labour Productivity Change.” Journal of Productivity Analysis 42: 35–43. Balk, B. M. 2015. “Measuring and Relating Aggregate and Subaggregate Total Factor Productivity Change Without Neoclassical Assumptions.” Statistica Neerlandica 69: 21–48. Balk, B. M. 2016. “The Dynamics of Productivity Change: A Review of the Bottom-Up Approach.” In Productivity and Efficiency Analysis, edited by W. H. Greene, L. Khalaf, R. C. Sickles, M. Veall and M.-C. Voia, 15–49. Proceedings in Business and Economics. Cham, Switzerland: Springer International. Balk, B. M. 2018. “Aggregate Productivity and Productivity of the Aggregate: Connecting the Bottom-Up and Top-Down Approaches.” In Productivity and Inequality, edited by W. H. Greene, L. Khalaf, P. Makdissi, R. C. Sickles, M. Veall and M.-C. Voia, 119–141. Proceedings in Business and Economics. Cham, Switzerland: Springer International.. Ball, V. E., R. Färe, S. Grosskopf, and D. Margaritis. 2015. “The Role of Energy Productivity in U.S. Agriculture.” Energy Economics 49: 460–471. Burns, A. F. 1930. “The Measurement of the Physical Volume of Production.” Quarterly Journal of Economics 44: 242–262. Calver, M., and A. Murray. 2016. “Decomposing Multifactor Productivity Growth in Canada by Industry and Province, 1997–2014.” International Productivity Monitor 31: 88–112. Center for the Study of Living Standards, Ontario, Canada. Dhyne, E., A. Petrin, V. Smeets, and F. Warzynski. 2014. “Import Competition, Productivity and Multi-Product Firms.” Working Paper Research No. 268 National Bank of Belgium, Brussels. Diewert, W. E. 1978. “Superlative Index Numbers and Consistency in Aggregation.” Econometrica 46: 883–900. Diewert, W. E. 1992. “The Measurement of Productivity.” Bulletin of Economic Research 44: 163–198. Diewert, W. E. 2014. “The Treatment of Financial Transactions in the SNA: A User Cost Approach.” Eurona 1: 73–89. Diewert, W. E. 2016. “Decompositions of Productivity Growth into Sectoral Effects: Some Puzzles Explained.” In Productivity and Efficiency Analysis, edited by W. H. Greene, L. Khalaf, R. C. Sickles, M. Veall, and M.-C. Voia, 1–14. Proceedings in Business and Economics. Cham, Switzerland: Springer International. Diewert, W. E., and K. J. Fox. 2010. “On Measuring the Contribution of Entering and Exiting Firms to Aggregate Productivity Growth.” In Price and Productivity Measurement, Vol. 6: Index Number Theory, edited by W. E. Diewert, B. M. Balk, D. Fixler, K. J. Fox, and A. O. Nakamura, 41–66. Vancouver: Trafford Press. www.vancouvervolumes.com, www.indexmeasures.com.

Empirical Productivity Indices and Indicators 119 Revised version of Discussion Paper No. 05–02, Department of Economics, University of British Columbia, Vancouver, 2005. Diewert, W. E., and A. O. Nakamura. 2003. “Index Number Concepts, Measures and Decompositions of Productivity Growth.” Journal of Productivity Analysis 19: 127–159. Dumagan, J. C., and B. M. Balk. 2016. “Dissecting Aggregate Output and Labour Productivity Change: A Postscript on the Role of Relative Prices.” Journal of Productivity Analysis 45: 117–119. Dumagan, J. C., and V. E. Ball. 2009. “Decomposing Growth in Revenues and Costs into Price, Quantity and Total Factor Productivity Contributions.” Applied Economics 41(23): 2943–2953. Felipe, J., and J. S. L. McCombie. 2013. The Aggregate Production Function and the Measurement of Technical Change: “Not Even Wrong.” Cheltenham, UK; Northhampton, MA: Edward Elgar. Gollop, F. M. 1979. “Accounting for Intermediate Input: The Link Between Sectoral and Aggregate Measures of Productivity Growth.” In Measurement and Interpretation of Productivity, edited by A. Rees, 318–333. Report of the Panel to Review Productivity Statistics, Committee on National Statistics, Assembly of Behavioral and Social Sciences, National Research Council. Washington, DC: National Academy of Sciences. Gordon, R. J. 2016. The Rise and Fall of American Growth: The U.S. Standard of Living since the Civil War. Princeton, NJ: Princeton University Press. Grifell-Tatjé, E., and C. A. K. Lovell. 2014. “Productivity, Price Recovery, Capacity Constraints and Their Financial Consequences.” Journal of Productivity Analysis 41: 3–17. Grifell-Tatjé, E., and C. A. K. Lovell. 2015. Productivity Accounting: The Economics of Business Performance. New York: Cambridge University Press. Griliches, Z., and H. Regev. 1995. “Firm Productivity in Israeli Industry, 1979–1988.” Journal of Econometrics 65: 175–203. Harrison, J. S., and A. C. Wicks. 2013. “Stakeholder Theory, Value and Firm Performance.” Business Ethics Quarterly 23: 97–124. Jorgenson, D. W., M. S. Ho, and K. J. Stiroh. 2005. Productivity, Vol. 3: Information Technology and the American Growth Resurgence. Cambridge, MA: MIT Press. Lawrence, D., W. E. Diewert, and K. J. Fox. 2006. “The Contributions of Productivity, Price Changes and Firm Size to Profitability.” Journal of Productivity Analysis 26: 1–13. Marr, B. 2012. Key Performance Indicators: The 75 Measures Every Manager Needs to Know. Harlow, UK: Pearson. Muntean, T. 2014. “Intangible Assets and Their Contribution to Labour Productivity Growth in Ontario.” International Productivity Monitor 27: 22–39. Center for the Study of Living Standards, Ontario, Canada. Murray, A. 2016. “Partial versus Total Factor Productivity Measures: An Assessment of their Strengths and Weaknesses.” International Productivity Monitor 31: 113–126. Center for the Study of Living Standards, Ontario, Canada. Schreyer, P. 2001. Measuring Productivity: Measurement of Aggregate and Industry-Level Productivity Growth. Paris: OECD. Vancauteren, M., E. Veldhuizen, and B. M. Balk. 2012. “Measures of Productivity Change: Which Outcome Do You Want?” Paper presented at the 32nd General Conference of the IARIW, Boston, MA, August 5–11. www.iariw.org.

Chapter 3

The US Bu re au of L ab or Stat i st i c s Productivit y Pro g ra m Lucy P. Eldridge, Chris Sparks, and Jay Stewart 1

3.1. Introduction The Bureau of Labor Statistics (BLS) at the US Department of Labor produces the official productivity statistics for the US economy.2 The BLS productivity measurement program has evolved over time, reflecting changes in available source data, improvements in methodology, and the demands of data users for new data products. The BLS’s studies of output per hour in individual industries began in the 1920s, and these early measures related output to the number of workers. The productivity program expanded to about 50 industries in the 1930s and 1940s with a focus on the manufacturing sector. Since then, the BLS industry productivity program has evolved from publishing occasional industry-specific studies to the regular publication of annual measures of labor productivity for manufacturing and service providing industries. Labor productivity data for the aggregate economy were first published in 1959, following the 1954 development of real gross national product (GNP) from the National Income and Product Accounts (NIPA) at the US Department of Commerce. The BLS began publishing quarterly measures of labor productivity growth in 1976. In 1983, the BLS developed and began publishing a broader measure of productivity, multifactor productivity (MFP), which is also referred to as total factor productivity (TFP).3 A major part of this effort was developing the methodology for estimating capital input from investment and other data. Today, the BLS continues to produce productivity statistics for major sectors of the US economy and individual industries. The BLS also publishes a number of related measures that shed light on the elements of productivity growth, including output and

122 Lucy P. Eldridge, Chris Sparks, and Jay Stewart implicit price deflators for output, employment, hours worked, unit labor costs, real and nominal hourly compensation, and unit non-labor payments. Labor’s share is another measure that can be constructed from these pieces, although it is not included in the productivity news releases. In constructing measures of labor and multifactor productivity, the BLS combines data from a number of sources, makes adjustments to ensure consistency, and removes known sources of bias specific to productivity measurement. The basic components of its productivity measures are output, labor inputs, capital inputs, and intermediate inputs. In the following, we describe how the BLS constructs these components and estimates productivity growth, and we discuss some of the measurement issues that the BLS faces going forward. The rest of this chapter is laid out as follows. In section 3.2, we describe the industry and sector coverage of the BLS’s productivity measures, and explain why certain sectors are excluded. Section 3.3 describes the growth accounting framework that the BLS uses to construct its labor and multifactor productivity measures. Section 3.4 describes the BLS’s output concepts. BLS methodologies for constructing labor, capital and other inputs are described in sections 3.5, 3.6, and 3.7. Section 3.8 discusses other measures that the BLS produces, and we wrap up the chapter with a look at ongoing projects in section 3.9.4

3.2. Industry and Sector Coverage Ideally, BLS productivity statistics would measure productivity for the US economy at the most aggregate level of domestic output, gross domestic product (GDP).5 However, the real output of general government, private households, and nonprofit institutions are estimated using data on labor compensation and other inputs, which implies little or no productivity growth for these sectors. The output of general government is measured as the sum of the compensation of general government employees and the consumption of general government fixed capital, which is a proxy for the capital services derived from general government assets. The output of the private household sector is measured as the compensation of paid employees of private households plus the rental value of owner-occupied housing; and the output of nonprofit institutions serving individuals is measured as the compensation paid to their employees plus the rental value of buildings and equipment owned and used by these institutions. The trends in output measured using compensation data will, by construction, move with inputs and thus will tend to imply little or no labor productivity growth.6 Therefore, to get a more accurate picture of productivity growth, the BLS focuses on the business sector, which excludes these activities and constitutes about 74% of GDP. The BLS produces quarterly and annual labor productivity measures for the business, nonfarm business, and nonfinancial corporate sectors.7 Excluding the farm sector, which is small in the United States, reduces the volatility of labor productivity.

The US Bureau of Labor Statistics Productivity Program 123 Average annual percent changes 4.0

3.5

3.0

2.5

2.0

1.5

1.0

0.5

0.0

1948 Q4– 1953 Q2– 1957 Q3– 1960 Q2– 1969 Q4– 1973 Q4– 1980 Q1– 1981 Q3– 1990 Q3– 2001 Q1– 2007 Q4– 1953 Q2 1957 Q3 1960 Q2 1969 Q4 1973 Q4 1980 Q1 1981 Q3 1990 Q3 2001 Q1 2007 Q4 2016 Q2

Figure 3.1. Labor productivity growth rates during business cycles since 1947. Source: US Bureau of Labor Statistics.

Output for the nonfinancial corporate sector is equal to that of the business sector less the output of financial corporations and unincorporated businesses.8 The BLS also produces quarterly and annual measures of labor productivity for manufacturing, and durable and nondurable manufacturing.9,10 The labor productivity series for the business and nonfarm business sectors extend from 1947 onward, for nonfinancial corporations from 1958 onward, and for manufacturing (total, durable, and nondurable sectors) from 1987 onward. Figure 3.1 shows productivity growth from the fourth quarter of 1948 through the second quarter of 2016. Here you can see the high productivity growth of the postwar period and the 1960s. These periods contrast with the low productivity growth of the 1970s and the even-lower growth since the Great Recession. The BLS also produces annual labor productivity measures for all three-and four-digit North American Industry Classification System (NAICS) industries in manufacturing, wholesale, and retail trade and for a variety of industries in mining and the services- providing sector.11 Figure 3.2 shows the distribution of industry productivity growth rates for the four-digit industries for which the BLS publishes estimates. Here we can see how the period after the Great Recession contrasts with earlier periods that also include recessions—productivity growth is negative for a large fraction of industries. In addition to labor productivity, the BLS produces annual MFP measures for the private business sector, the private nonfarm business sector, manufacturing sectors, all three-and four-digit NAICS manufacturing industries, roughly two-digit

0 2.1 to 4.0

0.1 to 2.0

10.1 or more

10

10.1 or more

20 8.1 to 10.0

30

8.1 to 10.0

40 6.1 to 8.0

50

6.1 to 8.0

60 4.1 to 6.0

2007–2014

4.1 to 6.0

2.1 to 4.0

0.1 to 2.0

70 –1.9 to 0.0

–3.9 to –2.0

–5.9 to –4.0

–7.9 to –6.0

–9.9 to –8.0

–10.0 or Less

Number of Industries 70

–1.9 to 0.0

–3.9 to –2.0

–5.9 to –4.0

–7.9 to –6.0

–9.9 to –8.0

–10.0 or Less

Number of Industries

10.1 or more

8.1 to 10.0

6.1 to 8.0

4.1 to 6.0

2.1 to 4.0

0.1 to 2.0

–1.9 to 0.0

–3.9 to –2.0

–5.9 to –4.0

–7.9 to –6.0

–9.9 to –8.0

–10.0 or Less

Number of Industries 70 1990–2000

60

50

40

30

20

10

0

2000–2007

60

50

40

30

20

10

0

Figure 3.2. Distribution of industry productivity for selected years.

Source: US Bureau of Labor Statistics.

The US Bureau of Labor Statistics Productivity Program 125 Table 3.1 Summary of BLS Productivity Measures and Coverage

Labor Productivity

Multifactor Productivity

Frequency

Sectors (Start Year)

Quarterly Annual

Business (1947) Nonfarm business (1947) Nonfinancial corporations (1958)

Annual

Manufacturing (1987) Durable (1987) Nondurable (1987)

Annual

Private business sector (1987) Private nonfarm business (1987) Manufacturing (1987)

Annual

Industries

Output Concept Value added

All three-and four-digit Sectoral NAICS industries in Output manufacturing, wholesale, and retail trade; selected industries in mining and service-producing (1987) Value added

All three-and four-digit Sectoral NAICS manufacturing Output industries and two service- sector industries (1987)

non-manufacturing industries,12 and the total economy, as well as for the three-digit air transportation industry and the six-digit line-haul railroads industry. The BLS restricts its coverage to the private business sector for its aggregate MFP measures because the appropriate capital data are not available for the government sector. In particular, subsidies account for a large fraction of government capital income, which makes it impossible to estimate the requisite rental prices and cost-share weights. The MFP measures for the private business and private nonfarm business sectors, and the total economy are calculated as value-added output per unit of combined labor and capital input, while the manufacturing and industry measures are calculated as sectoral output per unit of combined capital (K), labor (L), energy (E), materials (M), and purchased business services (S) inputs: KLEMS. Table 3.1 summarizes the different BLS productivity measures, and helps illustrate how they are related.

3.2.1. Preliminary Annual Estimates Major sector labor productivity measures are published shortly after the end of the reference quarter but, due to source data availability, there can be a considerable lag in the publication of MFP data and the industry-level labor productivity measures. To meet

126 Lucy P. Eldridge, Chris Sparks, and Jay Stewart data users’ demand for timely data, the BLS produces preliminary estimates based on a simplified methodology. For private business and private nonfarm business sector MFP, output and hours data are available, shortly after the end of the reference year. But the data required for the capital measures are not available until long after the end of the reference year. The simplified methodology uses fewer more-aggregated asset categories than the final published estimates, and a simpler methodology for estimating rental prices. Final estimates are published when complete data become available. Preliminary estimates of industry- level labor productivity growth for selected manufacturing, mining, and service industries are available within five months after the end of the reference period. As time passes, more and better data are available for generating preliminary estimates. Therefore, the BLS must balance the cost to users of delaying publication with the advantages of having better data to generate preliminary estimates.

3.3. Labor and Multifactor Productivity Quarterly labor productivity growth, which is designated by the US government as a Principal Federal Economic Indicator (PFEI), is calculated as the percentage growth in value-added output less the percentage growth in hours worked:

. Q VA H LP VA = VA − , H Q

(3.1)

where Q is output and H is hours, and the dot notation indicates the first derivative with respect to time. This is a simple and easy-to-calculate productivity measure, but it does not tell the whole story. To provide a more complete picture, the BLS publishes estimates of MFP growth, which also accounts for capital and changes in the composition of the labor force. The BLS estimates MFP growth using a growth accounting framework that assumes Hicks-neutral technical change and constant returns to scale (see Solow 1957):

. Q VA K VA L −s , MFP VA = VA − sVA K K L L Q

(3.2)

where K is capital input, L is labor input, and sVA and sVA are cost share weights. K L Labor input can be broken down into the growth in hours (H) and labor composition (LC), which accounts for changes in the demographic composition of the labor force. Expanding equation (3.2), we have:

The US Bureau of Labor Statistics Productivity Program 127

. . . VA Q VA VA K VA H LC Q VA VA K VA H VA LC (3.2ʹ ) MFP = VA − sK −s + = −s −s −s K L H LC QVA K K L H L LC Q

The MFP measure was originally created to explain labor productivity growth and, by rearranging terms in equation (3.2ʹ ), we can see the relationship between labor and multifactor productivity: . . VA Q VA H . VA VA K H VA LC LP = VA − = MFP + sK − + sL H LC K H Q

(3.3)

Thus, value-added labor productivity growth is equal to MFP growth plus the growth in the capital-to-labor ratio (capital intensity) weighted by capital’s share of cost plus the growth in labor composition weighted by labor’s share of cost. Table 3.2 shows how MFP growth has varied over time and its relationship to labor productivity. Over a given time period, the average logarithmic growth rate of capital intensity, labor composition, and MFP will sum to the average logarithmic growth rate

Table 3.2 Labor Productivity and the Contributions of Capital Intensity, Labor Composition, and Multifactor Productivity to Labor Productivity in the Private Nonfarm Business and Private Business Sectors for Selected Periods, 1987–2014 Compound Annual Growth Rates Private nonfarm business

1987– 1987– 1990– 1995– 2000– 2007– 2013– 2014 1990 1995 2000 2007 2014 2014

Labor productivity

2.0

1.6

1.6

2.9

2.6

1.2

0.7

Contribution of capital intensity

0.8

0.7

0.6

1.2

1.0

0.6

–0.1

Contribution of information- processing equipment

0.3

0.3

0.3

0.6

0.3

0.1

0.0

Contribution of research and development

0.1

0.1

0.1

0.1

0.1

0.1

0.0

Contribution of all other intellectual property products

0.2

0.2

0.4

0.4

–0.1

0.1

0.1

Contribution of all other capital services

0.2

0.1

–0.1

0.1

0.7

0.2

–0.3

Contribution of labor composition

0.3

0.2

0.5

0.2

0.2

0.3

0.1

Multifactor productivity

0.9

0.7

0.5

1.5

1.4

0.4

0.7

Contribution of R&D to multifactor productivity

0.2

0.2

0.2

0.2

0.2

0.1

0.1

128 Lucy P. Eldridge, Chris Sparks, and Jay Stewart of labor productivity. To illustrate, the strong growth in labor productivity in the 1995– 2000 and 2000–2007 periods can be mainly attributed to strong growth in capital intensity and MFP. The contribution of labor composition was relatively small, and not very different from other time periods.

3.4. Output Concepts and Industry Productivity Productivity estimates can be computed using any one of three output concepts: gross output, sectoral output, or value-added output.13 Gross output is the total output produced by an industry or sector, while value-added output is equal to gross output less all intermediate inputs used in production. Sectoral output lies between gross output and value-added output and is equal to gross output less only those intermediate inputs that are produced within that industry or sector; intermediate inputs used in production from outside the industry are not removed. Thus, sectoral output represents the value of output leaving the sector or industry.14 For detailed industries, sectoral output is very close to gross output because very few industry outputs are used as intermediate inputs in the same industry. Going from detailed industries to more-aggregated industries and major sectors, sectoral output converges toward value-added output. For the total economy, value-added and sectoral output are the same except for imported intermediate inputs.15 Different output concepts are useful for answering different questions.16 Value-added labor productivity more closely reflects an industry’s ability to translate labor hours into final income, while sectoral-output-based labor productivity measures the efficiency with which an industry transforms labor hours into output.17 The BLS uses value-added output for its major sector labor and multifactor productivity measures. For industry- level productivity measures, the BLS calculates both labor and multifactor productivity using sectoral output within a KLEMS framework so that the MFP measures can be used to explain labor productivity growth. Although the volume of intrasectoral transactions does not materially affect KLEMS MFP calculations, it does affect labor productivity because intermediate inputs that are produced within the industry would be double- counted if a gross output concept were used. Specifically, the intermediate input would show up twice in output—once as an intermediate output and once as part of the final output that is sold outside the industry—but the hours used to produce the intermediate input are counted only once. By using the sectoral output concept for both, it is easier to see the relationship between labor productivity and MFP. Starting with MFP, we have:

. LC SO E SO M . SO Q SO SO K SO H −i MFP = SO − sK s s − sSSO − − − sL + E M −i M −i K H LC E Q

S , S

(3.4)

The US Bureau of Labor Statistics Productivity Program 129 where E denotes energy, M−i denotes materials purchased from outside industry i, and S denotes purchased services. This is the standard KLEMS model, but materials exclude those that are produced in the industry. Rearranging equation (3.4), sectoral-output- based labor productivity is:

. Q SO H LP SO = SO − H Q . M E H K H H = MFP SO + sKSO − + s ESO − + s MSO− i − i − E H K H M −i H . S H LC + sSSO − + sLSO LC S H

(3.5)

Under both output concepts, value-added and sectoral, changes in labor productivity can be due to technological progress, increased capital intensity, greater economies of scale, improved management techniques, and changes in the skills of the labor force.18 The key difference between the two concepts is that sectoral-output-based labor productivity includes the effects of substituting other inputs (EMS) for labor, whereas value-added labor productivity does not.19 Sectoral-output-based labor productivity will therefore grow with increased outsourcing of labor and purchases of intermediate inputs. But if these purchased intermediate inputs are subtracted from the value of output, as with value- added output, they can no longer be a source of productivity growth.20 Thus, outsourcing of labor has a smaller effect on industry-level value-added labor productivity, because the substitution of purchased services for labor reduces both output and labor input.21

3.4.1. Output Data The BLS gets its output data from the US Bureau of Economic Analysis (BEA) and the US Census Bureau. The real value-added output used for business-sector productivity measures come from the BEA’s NIPAs. BEA calculates business-sector output (business, nonfarm business, private business, and private nonfarm business) by removing from GDP the gross product of general government, private households, and nonprofit institutions. Output measures for the manufacturing sectors and industries begin with value of industry shipments from the US Census Bureau. These gross output data are then adjusted to remove transactions that occur between establishments within the sector or industry, and to account for resales and changes in inventories. Intrasectoral transactions are calculated using delivered-costs data from the US Census Bureau’s Materials Consumed by Kind table, which includes detailed data on the products that are purchased by industries for use in the manufacturing process. The BLS adjusts the delivered costs using BEA input–output tables to remove the difference between the price the producer receives and the price that the consuming establishment pays. BEA Import Matrixes are

130 Lucy P. Eldridge, Chris Sparks, and Jay Stewart then used to remove imported materials so that only domestically produced products are included in the intrasectoral adjustment. In years where the Census does not publish the Materials Consumed by Kind table, the BLS uses value-of-shipments and cost- of-materials data from the Annual Survey of Manufactures to estimate intrasectoral transactions. The BLS produces labor productivity measures for non-manufacturing industries only if output can be defined and measured in a way that is independent from labor input measures. For those non-manufacturing industries, the output measures are constructed primarily using data from the economic censuses and annual surveys of the US Census Bureau together with information on price changes from the BLS.22 Real output is most often derived by deflating nominal sales or values of production using BLS price indexes and removing intra-industry transactions, but for a few industries, such as education, commercial banking, line-haul railroads, and industries in the mining sector, it is measured by “physical” quantities of output.23 For NIPA- level non-manufacturing industries, for which the BLS publishes MFP measures, sectoral output is estimated using BEA gross output estimates and intrasectoral transaction data based on the BEA input–output use tables—specifically, the use tables before redefinitions.24 In order to publish quarterly estimates of labor productivity for manufacturing, the BLS estimates sectoral output by adjusting annual manufacturing output data using a quarterly reference series and a quadratic minimization formula.25 The quarterly reference series is constructed from the Federal Reserve Board monthly Indexes of Industrial Production. Due to a lag in the availability of the annual benchmark data, quarterly and annual manufacturing output measures are estimated by extrapolation based on the changes in the Indexes of Industrial Production.

3.5. Labor Input For labor productivity, labor input is measured as total hours worked. This is a straightforward concept, but it treats every hour the same, regardless of the labor force’s experience and education. For MFP, a more comprehensive measure of labor input is used. Specifically, the MFP labor input measure accounts for changes in the composition of the labor force in a way that gives greater weight to more highly paid (and presumably more experienced and higher skilled) labor hours. In the following, we describe how the BLS estimates hours worked and calculates its labor composition index.26

3.5.1. Hours Worked Total hours worked includes all time available for production and excludes paid vacations and other forms of leave, which are viewed as benefits.27 This definition of

The US Bureau of Labor Statistics Productivity Program 131 hours worked is consistent with the International Labour Organization–United Nations (ILO-UN) resolution on measuring working time. The BLS productivity program uses data from three BLS surveys to construct its estimate of work hours: the Current Employment Statistics survey (CES), the Current Population Survey (CPS), and the National Compensation Survey (NCS).28 The CES is a monthly payroll survey that collects data on employment and hours paid from a sample of about 623,000 worksites.29 The CPS is a monthly survey of approximately 60,000 households that collects data on employment status, hours worked, and demographic characteristics.30 The NCS is a quarterly survey of 28,900 occupations from a sample of 6,900 establishments that collects data on hours worked and paid leave.31 The BLS uses the CES as its main source of hours data, because industry classifications in establishment surveys are more consistent with those used in the surveys that collect output data and because the CES’s large sample size permits considerable industry detail. However, the CES hours data are hours paid and, until 2006, covered only production (in goods-producing industries) and nonsupervisory (in services-providing industries) employees.32,33 Hereafter, we will refer to these workers as production workers. In contrast, the CPS collects data on hours worked by all workers, but does not use industry classifications that are completely consistent with those in the output data sources.34 Moreover, the small sample size of the CPS limits the amount of industry detail. The CPS is used to estimate hours for nonproduction workers. The NCS is used to convert CES production worker hours to an hours-worked basis. The BLS begins by taking published estimates of average weekly hours paid by production workers from the CES, and adjusting them from an hours-paid to an hours- worked basis using a ratio estimated from the NCS.35 This adjustment ensures that changes in vacation, holiday, and sick pay do not affect hours growth. More formally, hours worked by production workers is given by:

CES NCS CES HOURSPW = AWH PW × hwhpPW × EMPPW × 52

(3.6)

where AWH is average weekly hours, hwhp is the hours-worked-to-hours-paid ratio, and EMP is employment. Superscripts indicate the survey, and subscripts indicate the worker category. Note that production worker hours, as calculated in equation (3.6), do not capture off-the-clock hours. Off-the-clock work is not an issue for hourly paid workers, who comprise about 70% of production workers. But it may be an issue for the 30% of production workers who are not paid hourly. Employers are not required to keep records for these workers and, for this reason, CES respondents are instructed to report standard workweeks. This means that changes in off-the-clock hours worked by non-hourly-paid production workers are not captured. To estimate nonproduction worker hours, the BLS constructs a ratio of nonproduction worker average weekly hours to production worker average weekly hours using data from the CPS and then multiplies the NCS-adjusted production worker hours by that ratio.36 Note that this CPS-ratio adjustment does capture off-the-clock work for

132 Lucy P. Eldridge, Chris Sparks, and Jay Stewart nonproduction workers. Hours worked by nonproduction workers are calculated as follows:

(

)

CES NCS HOURSNPW = AWH PW × hwhpPW ×

CPS AWH NPW CES × EMPNPW × 52 CPS AWH PW

(3.7)

Total hours worked for all employees is equal to the sum of production worker hours and nonproduction worker hours.37 Total hours for all workers are equal to total employee hours plus the hours worked by farm workers, the unincorporated self-employed, and unpaid family workers, all of which are estimated using hours-worked data from the CPS.38

3.5.1.1. Hours Research at the Bureau of Labor Statistics It is worth noting that the BLS differs from statistical agencies in many other countries in that it builds its estimates of hours worked from three different data sources, with data from the monthly establishment survey being the primary source. It has been documented that the two primary BLS hours series differ in level and trend (Abraham, Spletzer, and Stewart 1998, 1999; Frazis and Stewart 2010). The BLS has conducted a significant amount of research to evaluate the quality of its hours data.39 Several studies examine the quality of hours reports from the CPS. Frazis and Stewart (2004, 2009, 2010) examined the claim that respondents in household surveys such as the CPS tend to overstate their hours worked. They compared hours data from the CPS to hours data from the American Time Use Survey (ATUS) and found that hours are correctly stated on average, although some groups overstate hours while other groups understate hours. They also found that multiple job holding is understated in the CPS and hours worked on second jobs are overstated (Frazis and Stewart, 2009). The net effect is that total hours worked on second jobs is slightly understated in the CPS. They also found that people tend to work more hours during CPS reference weeks compared with nonreference weeks.40 Eldridge and Pabilonia (2010) use data from the CPS and the ATUS to examine whether off-the-clock work done at home is counted. They find that workers who bring work home tend to work less at the office, but work more hours overall. They find that there may be a small downward bias (less than 1%) in the official level of hours worked, but this does not affect growth rates. Frazis and Stewart (2010) investigated the differences in level and trend between the CPS and the CES. Their approach was to simulate the CES data using data from the CPS. They found that virtually all of the difference in level could be accounted for by differences in coverage (all workers in the CPS vs. production/nonsupervisory workers in the CES), the treatment of multiple job holders (counted once in the CPS vs. multiple times in the CES), and the hours concept (hours worked in the CPS vs. hours paid in the CES). The difference in trends was harder to reconcile. Frazis and Stewart investigated

The US Bureau of Labor Statistics Productivity Program 133 a number of possible explanations but were only able to explain about one-third to one-half of the divergence in trend. They did, however, identify three industries that accounted for a large portion of the divergence: retail trade, leisure and hospitality, and business and professional services.

3.5.2. Labor Composition The labor composition term in the MFP equation (3.2ʹ ) accounts for changes in the demographic composition of the labor force and is a rough proxy for labor quality.41 The BLS labor composition index accounts for the effect of shifts in the age, education, and gender composition of the workforce on the efficiency of labor. The demographic information required for labor composition calculations are available only in household surveys, such as the CPS. The BLS uses the CPS Outgoing Rotation Group (ORG) data, which has information on workers’ hours and earnings, as well as a wealth of demographic information. In addition, ORG data files have a large enough sample to allow the labor composition adjustment to be calculated at the three- digit industry level.42,43 The BLS divides the CPS sample into gender × age category × education level worker groups, and computes the year-to-year growth in hours for each group. These growth rates are Törnqvist aggregated into a measure of total labor input growth, where the weight for each group is its share of total labor compensation:44

∆Labor Inputt = ∑ j

1 s +s × ∆HOURS jt , 2 jt j ,t −1

(

)

(3.8)

where

s jt =

w jt × HOURS jt

, ∑ jw jt × HOURS jt

(3.9)

and w jt is the median wage of workers in worker group j in year t. Theoretically, the mean wage is more appropriate, but the BLS uses medians because some groups have only a few observations, and medians are not as sensitive to outliers. The change in labor composition is calculated as follows:

∆Labor Compositiont = ∆Labor Inputt − ∆Total Hourst

(3.10)

Note that the second term, the (unweighted) change in total hours, is also calculated using the CPS ORG data to be consistent with the data used to calculate labor input. The

134 Lucy P. Eldridge, Chris Sparks, and Jay Stewart labor composition index is calculated by chaining the annual growth rates back to the base period.

3.6. Capital Inputs The BLS estimates of capital input cover a wide range of asset types, including fixed business equipment, structures, artistic originals, research and development, own- account software, inventories, and land.45 Financial assets are excluded from capital input measures, as are owner-occupied residential structures. The BLS, like other statistical agencies, estimates capital inputs by first estimating the productive capital stock using the perpetual inventory method and then estimating the rental price of capital.46 The BLS calculates capital stocks and rental prices for 90 asset types in 60 NIPA-level NAICS industries (and 29 assets for 86 4-digit manufacturing industries).

3.6.1. The Productive Capital Stock The productive capital stock is measured as the sum of past investments net of deterioration. Past investments are assumed to decline in productive capacity according to an age-efficiency function, λ(a,Ω)—where a is the age of the asset and Ω is its maximum service life—that represents the proportion of the investment’s original productive capacity that remains at age a. Letting Kt denote the (net) productive capital stock of an asset in year t and It–a denote investment expenditures in year t–a (that is, t–a is the vintage of the asset), the productive capital stock for a group of identical assets that have a maximum service life of Ω is given by:47 Ω

K t = ∑λ (a, Ω ) It − a

(3.11)

a=0

The BLS assumes a hyperbolic age-efficiency function, which is given by:

(Ω − a ) (Ω − βa )

λ (a , Ω ) =

λ (a , Ω ) = 0

if a < Ω

(3.12)

if a ≥ Ω

where β ≤ 1 is a shape parameter and λ(⋅) is concave, linear, or convex in age, depending on whether β ⪌ 0 (β = 1 implies one-hoss shay48). The BLS assumes β = 0.75 for structures and β = 0.5 for equipment.49

The US Bureau of Labor Statistics Productivity Program 135 The age-efficiency function accounts for three avenues by which the productive capacity of a group of assets can decline. First, assets become physically less productive when used or require more “down time” for maintenance or repairs. The second avenue is through obsolescence. The third is through failure. With all three of these avenues, one would expect the age-efficiency function for an individual asset to be concave with respect to age. Thus, even a group of one-hoss-shay assets would have a concave age- efficiency function due to failure and obsolescence. The age-efficiency function in equation (3.12) assumes that all assets in the group have identical maximum service lives. However, the investment data that the BLS receive from BEA are for asset categories that contain assets that are similar, but likely have different maximum service lives.50 Within an asset category, maximum service lives can differ because the assets are heterogeneous or because the assets are used differently. Therefore, within an asset category, the BLS assumes a distribution of maximum serv ice lives for each type of asset and computes a cohort age-efficiency function that is a weighted average of the deterioration functions of the individual asset types.51 The cohort age-efficiency function for an asset category with mean service life of Ω is given by

λ ( a, Ω ) =

Ωmax

∫ φ (a, k ) ⋅ λ (a, k ) dk

(3.13)

Ωmin

where the limits of the integral are the upper and lower bounds of the distribution of service lives, and Ωmin = 0.02Ω and Ωmax = 1.98Ω . The BLS assumes that φ (⋅) is a modified truncated normal distribution with mean Ω and σ = 0.49Ω. It is derived by truncating the normal distribution at ±2 standard deviations ( Ω ± 0.98Ω ), shifting the density function downward so that it equals zero at the upper and lower bounds of the distribution, and then inflating the density function proportionately so that the final modified density, φ (⋅), integrates to 1. The final capital stock for each industry × asset-category combination is estimated by substituting equation (3.13) for the age-efficiency function in equation (3.11):

Kt =

Ωmax

∑ λ ( a, Ω ) I a=0

t −a

(3.14 )

Note that the upper limit of the summation differs from that in equation (3.11). In equation (3.11), maximum service life refers to homogeneous assets with a maximum service life of Ω. But in equation (3.14), the maximum service life is the maximum of the longest-lived asset in the asset category, Ωmax .

3.6.2. The Rental Price of Capital Conceptually, the rental price, or user cost, of capital is the opportunity cost of holding and using it for a period of time. The equilibrium rental price is equal to the forgone

136 Lucy P. Eldridge, Chris Sparks, and Jay Stewart earnings of money invested in the asset plus the decline in the asset’s value. In its simplest form, the rental price for a period is set to the price of the asset multiplied by the sum of the depreciation rate and the appropriate rate of return: ct = pt (rt + dt ) ,

(3.15)

where ct is the rental price for period t, pt is the price of the asset, rt is the rate of return, and dt is the rate of depreciation. The equation that the BLS uses to construct rental prices of capital is somewhat more complicated because it also accounts for inflation in the price of new assets and the effects of taxes. The BLS equation is given by: cijt =

(1 − u z

t t

(

− et ) Pj ,t −1rit + Pj ,t −1dijt − ∆Pj ,t −1 1 − ut

)+P

x

j ,t −1 t

(3.16)

where: ut is the corporate income tax rate; zt is the present value of $1 of the depreciation deduction; et is the effective rate of the investment tax credit (zero since 1979); rt is the nominal industry-specific internal rate of return on capital; dijt is the average rate of economic depreciation of asset j in industry i; Pj ,t −1 is the asset-specific (for asset category j) deflator for new capital goods; ∆Pj ,t −1 is the revaluation of assets due to inflation in new goods prices (asset category j); xt is the rate of indirect taxes. Rental prices are computed separately for each asset category × industry combination. Data are available, or easily calculated, for all of the variables in equation (3.16) except the internal rate of return, r.52

3.6.2.1. Wealth Stock and Depreciation The BLS calculates the wealth stock in order to estimate depreciation. The wealth stock differs from the productive capital stock in that it measures the financial value of the capital stock, rather than its productive capacity. Analogously, depreciation differs from deterioration in that it measures the decline in the financial value of the capital stock, rather than the decline in its productive capacity. However, the two measures are linked through the age-efficiency function. The main difference between the productive capital stock and the wealth stock is that the productive stock is a point-in-time measure (the current productive capacity of past investments), whereas the wealth stock is a forward- looking measure (the discounted value of the remaining productive capacity of the productive capital stock).

The US Bureau of Labor Statistics Productivity Program 137 The wealth stock is equal to the age-adjusted price of the asset multiplied by real investment summed over all asset cohorts that have positive productive capacity at time t: 2t

(

)

Wt = ∑ p τ − t , Ω ⋅ I 2t − τ

τ =t

(

(3.17)

)

where p τ − t , Ω is the price of a (τ–t)-year-old asset (group) that has an average maximum service life of Ω. The age-price function is derived from the cohort age-efficiency function and is equal to the discounted value of the remaining productive capacity of the asset divided by the discounted value of the productive capacity of the asset over its entire service life: 2 L +1

∑ p ( a, Ω ) = ∑

λ (α, Ω ) ⋅ (1 − ρ)

α −a

α =a 2 L +1

λ (α, Ω ) (1 − ρ)

α

α =0

(3.18)

where λ(a, Ω) is the cohort age-efficiency function (3.13), and ρ is the real discount rate, which is assumed to be 4% per year. The age-price function declines over time from 1, when the asset is new, to 0 at the end of its service life. It declines more quickly than the age-efficiency function, because the financial value of an asset declines—due to decreased remaining service life—even if its productive capacity does not.

3.6.2.2. The Internal Rate of Return The internal rate of return is calculated by setting capital income (which is available from BEA) equal to the product of the capital stock (from equation (3.14) and the rental price of capital (defined in equation (3.16)): (3.19)

Yit ≡ K it cit ,

where Yit is nominal capital income in industry i and Kit is the productive capital stock in industry i. Substituting equation (3.16) for ct and arranging terms, the internal rate of return is:

rit =

(

)

Yit − xK it Pi ,t −1 − K it Pi ,t −1dit − ∆Pi ,t −1 (1 − ut zt − et ) / (1 − ut ) K it Pi ,t −1 (1 − ut zt − et ) / (1 − ut )

(3.20)

where Pi,t–1 is the industry level price deflator:

Pi ,t −1 = ∑ j ∈J i

∑

K ij ,t −1

⋅

IijN,t −1

K ij ,t −1 I j ∈J i

R ij ,t −1

=∑

where Ji is the set of asset categories in industry i.53

j ∈J i

K ij ,t −1 K i ,t −1

Pj ,t −1

(3.21)

138 Lucy P. Eldridge, Chris Sparks, and Jay Stewart In about one-third of industries the BLS uses an external rate of return, which is calculated as the capital-stock-weighted average of the industry internal rates of return (including the negative rates). The external rate of return is used: when the internal rate is negative, when the change in the rental price is abnormally large, when the capital composition adjustment is abnormally large, or when capital income is negative. Negative internal rates of return are the most common reason for using the external rate. Note that capital income is available only for the corporate sector. For the non- corporate sector, the published data are for proprietors’ income, which includes capital income and proprietors’ labor compensation. The BLS estimates non-corporate capital income using data on proprietors’ income, average wages of wage and salary workers, and the non-corporate capital stock. The BLS computes independent estimates of proprietors’ labor compensation and capital income, sums these two components, and then inflates or deflates this sum so that it equals proprietors’ income. These independent estimates are generated by assuming that proprietors earn the same hourly wage as wage and salary workers (for labor compensation) and that non-corporate capital earns the same rate of return as corporate capital (for capital income).

3.6.2.3. Capital Input and Capital Composition The BLS assumes that the growth in capital inputs is approximately proportional to the growth of the productive capital stock. Capital input growth is calculated as a Törnqvist index, similar to MFP labor input as described in equation (3.8)—except that it is the Törnqvist aggregation of productive capital stock growth rates weighted by each asset × industry combination’s share of total capital cost (averaged over both years).54 The contribution of capital inputs to output growth can be decomposed into the contributions of changes in the capital stock and changes in the composition (“quality”) of the capital stock. The growth in capital input is calculated as described above. The effect of capital composition on output growth is simply the growth in capital input minus the growth in the (unweighted) capital stock. This decomposition is exactly analogous to the decomposition of labor input into changes in hours worked and changes in the composition of the workforce.55

3.7. Energy, Materials, and Purchased Business Services As the focus narrows to more specific industries that use sectoral output instead of value- added output, intermediate inputs (energy, materials, and purchased business services) take on a more visible role in productivity measurement and analysis. Manufacturing industry nominal values of materials, fuels, and electricity, as well as quantities of

The US Bureau of Labor Statistics Productivity Program 139 electricity consumed are obtained from the Annual Survey of Manufacturers and the economic censuses of the Census Bureau. Purchased business services are estimated using benchmark input–output tables and other annual industry data from the BEA. Prices of material and service inputs are based on the BLS price programs, the NIPAs, and the US Department of Agriculture. Prices of energy for manufacturing industries are obtained from economic censuses and annual surveys of the Census Bureau, as well as data from the Manufacturing Energy Consumption Survey of the Energy Information Administration, US Department of Energy. For NIPA- level non- manufacturing industries, intermediate inputs (energy, materials, and purchased business services) are obtained from BEA’s annual industry accounts. For air transportation, nominal values of cost of materials, services, fuels, and electricity are obtained from the Bureau of Transportation Statistics, US Department of Transportation, and are deflated with cost indexes from Airlines for America. For line-haul railroads, nominal values and price indexes are obtained from the Surface Transportation Board of the US Department of Transportation, and are supplemented with data from AMTRAK and other sources. Total intermediate inputs are calculated as a Törnqvist aggregation of these three input classes. In addition, purchased materials and services are adjusted to exclude transactions between establishments in the same industry or sector to maintain consistency with the sectoral output concept.

3.8. Other Measures In addition to the productivity measures described in the preceding, the BLS produces several related measures: labor compensation, unit labor costs, and labor share. These are not productivity measures per se, but they shed light on the BLS’s productivity statistics by showing the relationship between productivity and compensation.

3.8.1. Labor Compensation The BLS real hourly compensation data for the aggregate sectors of the US economy are based on data published by BEA as part of the national income accounts. These quarterly data include direct payments to labor—wages and salaries (including executive compensation), commissions, tips, bonuses, and payments in kind representing income to the recipients—and supplements to these direct payments. Supplements consist of vacation and holiday pay, all other types of paid leave, employer contributions to funds for social insurance, private pension and health and welfare plans, compensation for injuries, and so on.56 The BEA’s source data come from the BLS and the Census Bureau. For service- providing, mining, and utility industries, labor compensation is derived using annual

140 Lucy P. Eldridge, Chris Sparks, and Jay Stewart wage data from the Quarterly Census of Employment and Wages (QCEW) published by the BLS, along with data on employer costs for fringe benefits from the Census Bureau and the BEA. For manufacturing industries, annual payroll and fringe benefit data from the Census Bureau are used. The BEA compensation measures cover only wage and salary workers, which means that the BLS must estimate the labor compensation of proprietors. Hourly compensation for proprietors is assumed to be the same as that of the average employee in that sector; this measure is multiplied by the hours worked by proprietors in each sector to arrive at proprietors’ total labor compensation. The BLS productivity program’s labor compensation series is one of the most comprehensive wage series covering the private business sector. All private business sector workers are covered, and the measure includes virtually every form of compensation. Compared to other BLS earnings series (the production/nonsupervisory series from the Current Employment Statistics survey and an all-worker series from the Current Population Survey), the Labor Compensation series exhibits significantly stronger growth over the last 50 or so years, with the divergence becoming even greater starting in the late 1990s. Champagne, Kurmann, and Stewart (2015) compare these series and find that the main reason for the faster growth of the labor compensation series is the inclusion of non-wage compensation, which has grown in importance over the last few decades.

3.8.2. Unit Labor Costs Unit labor costs measure the cost of labor required to produce one unit of output. This measure can be derived by dividing an index of nominal labor compensation by an index of real industry output, or directly by dividing nominal compensation per hour by real output per hour. As such, unit labor cost measures describe the relationship between hourly compensation and labor productivity, and can be used as an indicator of inflationary pressure on producers. Increases in hourly compensation increase unit labor costs, while increases in labor productivity increases lower unit labor costs.

3.8.3. Labor Share Labor share is the portion of output that goes to labor, and is defined as nominal compensation divided by nominal output. Labor share relates the output of the private business sector to the compensation received by all workers in the private business sector, and tells us whether productivity growth has translated into higher incomes for labor. It also gives an indication of how competitive labor markets are—in a competitive labor market we would expect wages to rise with productivity. The decline in labor

The US Bureau of Labor Statistics Productivity Program 141 share over the last couple of decades has received a lot of attention in the popular press. The BLS does not publish data on labor share in its quarterly news release, although it is easily calculated from the data that are published.57 A paper by Elsby, Hobijn, and Şahin (2013) has taken issue with the BLS’s labor share measure, specifically how the BLS estimates proprietors’ labor compensation. As noted earlier, proprietors’ labor compensation is estimated under the assumption that proprietors earn the same wage as all wage and salary workers, which can result in proprietors’ labor compensation exceeding proprietors’ income, implying a negative return to capital (this happened in the 1980s—see Elsby et al. 2013). Elsby et al. prefer the approach to estimating proprietors’ labor compensation used in the BLS’s MFP statistics (described earlier). Unfortunately, the capital data needed to estimate proprietors’ capital income are available only with a substantial lag. The BLS is looking into ways of developing preliminary estimates of proprietors’ capital income that can be used until the final data become available.

3.8.3.1. The Compensation Gap Related to labor share is the compensation gap, which is equal to the difference between real output per hour (labor productivity) and real hourly compensation. This measure differs from labor share in that hourly compensation is deflated using the Consumer Price Index, while output per hour is deflated using an output price index. Thus, some of the gap is due to differences in the price indexes used. Figure 3.3 shows the growth rates of real hourly compensation, labor productivity, and the wage gap. The compensation gap has been growing since the 1960s, although there has been significant variation in the growth rates over time.58 Average annual percent changes 4.0

3.0

2.0

1.0

0.0

1948 Q4– 1953 Q2

1953 Q2– 1957 Q3

1957 Q3– 1960 Q2

1960 Q2– 1969 Q4

1969 Q4– 1973 Q4

1973 Q4– 1980 Q1

1980 Q1– 1981 Q3

1981 Q3– 1990 Q3

1990 Q3– 2001 Q1

2001 Q1– 2007 Q4

2007 Q4– 2016 Q1

Average Cycle

–1.0

–2.0 Compensation Gap

Labor Productivity

Real Hourly Compensation

Figure 3.3. Compensation gap, labor productivity, and real hourly compensation. Source: US Bureau of Labor Statistics.

142 Lucy P. Eldridge, Chris Sparks, and Jay Stewart

3.9. Ongoing Efforts The BLS is continually seeking ways to improve the quality of its existing data products by incorporating newly available data sources and developing improvements in methodology. And to meet ever-increasing customer demands for yet more timely data, the BLS is working to expand its data products and to speed up the production of existing products. Currently, the BLS is working on several projects to improve and expand its productivity measures. We describe a few of them in the following.

3.9.1. Improved Data on Asset Service Lives The BLS currently receives data on depreciation rates and service lives from the BEA.59 The BEA’s data on depreciation rates come from a variety of sources, which include studies by Hulten and Wykoff (1981a, 1981b), studies from the 1970s by the Office of Industrial Economics of the US Department of the US Treasury, and studies from the 1980s and 1990s by the Office of Tax Analysis of the Department of the Treasury.60 It is clear from the dates of these sources that the service lives used in BEA and BLS calculations are based on very old data. The BLS is currently exploring ways of improving data on asset service lives by first looking for existing data, such as data from trade associations and surveys conducted by other countries. The most promising of these, from the BLS’s perspective, is the survey conducted by Statistics Canada. The main difficulty with using these data is that the asset categories are not the same. In some cases the Statistics Canada classifications are more detailed, and in others they are more aggregated. The BLS has research underway to develop and evaluate new estimates of capital asset service lives using data from Statistics Canada.

3.9.2. Establishment-Level Productivity Statistics The BLS currently publishes estimates of productivity growth at a fairly detailed industry level, which allows one to examine the contributions of individual industries to aggregate productivity growth. Because the BLS uses aggregate, industry-level data, it is not possible to look beyond detailed industries. However, there is a large literature that shows that examining within-industry variation can provide important insights into aggregate productivity growth. (see Syverson 2011 for a nice summary). These studies have found that there is significant intra-industry dispersion in productivity (both labor productivity and MFP), and that high-productivity establishments are more likely to survive and grow compared with low-productivity establishments. A significant portion of aggregate productivity growth can be attributed to the reallocation of resources through entry and exit, but also among existing establishments as a result of

The US Bureau of Labor Statistics Productivity Program 143 resources being directed away from low-productivity establishments toward those with higher productivity.61 The BLS is currently working with the Census Bureau’s Center for Economic Studies on a project to produce two new data products: a data set that researchers can use to study productivity at the establishment level, and a periodic news release that examines the dispersion of productivity across establishments within various manufacturing industries. The project’s initial findings, which are summarized in Cunningham et al. (2018), indicate that, in addition to considerable dispersion within industries, there is also considerable “dispersion in dispersion.” That is, there is considerable variation, both between industries and within industries over time, in the amount of within-industry dispersion. Cunningham et al, also find that micro-level dynamics can account for a significant portion of industry productivity growth in official productivity statistics.

3.9.3. Output Concepts Multifactor productivity can be measured using one of three different output concepts: gross output, sectoral output, or value-added output. As noted earlier, gross output is the total output produced by an industry or sector, while value-added output is equal to gross output less all intermediate inputs used in production. The BLS uses value-added output in constructing productivity measures for major sectors of the US economy. However, for industry productivity measurement the BLS prefers sectoral output.62 To illustrate the impact of output and input choices on productivity measurement, BLS research is underway to estimate multifactor productivity and industry contributions to aggregate multifactor productivity using these three different, but related, definitions of output. This study will examine the relationship between MFP growth rates calculated using the three output concepts, as well as the different conclusions about industry contributions to aggregate productivity implied by each framework.

3.9.4. Expanded Industry Coverage As noted earlier, the BLS productivity program produces estimates for detailed industries. However, there is significantly less coverage of detailed industries in the services sector due to difficulties defining output in a way that is independent from measures of labor input (see Table 3.1).63 For detailed services industries, the BLS uses deflated revenues when they are available, and uses direct measures of output when good revenue data and deflators are not available. The BLS is continually undertaking research to expand coverage in hard-to-measure industries and most recently introduced the following new series: Construction: the BLS began publishing productivity measures for four construction industries in 2017. These measures use deflated revenues as the output measure, but they

144 Lucy P. Eldridge, Chris Sparks, and Jay Stewart take advantage of newly-available price deflators that are a significant improvement over those that were previously available. The methodology and results are described in Sveikauskas, Rowe, Mildenberger, Price, and Young (2015, 2016, and 2018). Education: the BLS introduced productivity measures for primary and secondary schools in 2017. For this industry, output cannot be measured using deflated revenues, because most schools are public.64 For this reason, the BLS measures output as school enrollment adjusted for quality using standardized test scores (Powers 2016a). The BLS has also constructed some experimental series that account for other types of school output, such as transportation and food services, and that account for the difference between disabled and nondisabled students (Powers 2016b). Commercial banking: the BLS revised its output and productivity measures for the commercial banking industry in 2012 (see Royster 2012). The previous BLS measure counted the different transactions, and aggregated them into a total output measure using employment-based weights. The new measure is a hybrid that uses a combination of deflated revenues (for fee-based services) and “physical” counts (for non-fee-based services). The new measure accounts for a wider range of bank services and uses revenue- based share weights (rather than the employment-based weights). This new measure better captures the broad range of services provided by the commercial banking industry.

3.9.5. Improved Measure of Proprietors’ Labor Compensation As noted earlier, the BLS assumes that proprietors earn the same hourly wage as wage- and-salary workers. The main drawback to this approach is that it treats the proprietors’ return to capital as a residual, which can be negative. The BLS is conducting research to address this issue. Because the data on proprietors’ capital income are not available until long after the end of the reference period, BLS would need to generate preliminary estimates of proprietor’s labor compensation using data from prior periods. As with other BLS data products, the estimates would be revised as new data become available, and final estimates would not be available until the official MFP (and capital) data are published.

3.9.6. Improved Estimates of Nonproduction and Supervisory Hours The BLS introduced the CES all-employee hours series in 2006. However, the fact that the CES form instructs respondents to report the standard workweek for salaried employees is problematic for productivity measurement. Hourly workers are paid for every hour they work, that is not necessarily the case for salaried workers. Thus, the CES hours data may miss some cyclical variation in hours worked by salaried workers.65

The US Bureau of Labor Statistics Productivity Program 145 The BLS’s nonproduction-to-production worker ratio in equation (3.7) captures some of this variation. But, it implicitly assumes that production/nonsupervisory workers in the CES are paid for every hour they work. In fact, analysis of CPS data indicates that about 30% of production workers are not paid by the hour. The BLS is researching the possibility of adjusting the production-worker hours to account for off-the-clock work by non-hourly workers. As part of this research, BLS is looking into the possibility of using the CES all-employee hours data (adjusted to account for paid time off and off-the-clock work).

Notes 1. Disclaimer: Any views expressed in this chapter are ours and do not necessarily reflect those of the BLS. 2. The United States has a decentralized statistical system that consists a number of statistical agencies housed in different executive departments of the US government, and the Federal Reserve Board. The Bureau of Labor Statistics is housed in the Department of Labor, and produces data related to the labor force, including employment, unemployment, price, compensation, and productivity statistics. Much of this introduction draws from Dean and Harper (2001), which is a detailed description of the evolution of the BLS productivity program. BLS productivity data can be found at http://www.bls.gov/bls/ productivity.htm 3. MFP is a broader measure of productivity in which an index of output is divided by an index of combined inputs—capital, labor, and, for more detailed industries, energy, materials, and purchased services. 4. For more information on the BLS methods, see “Technical Information about the BLS Major Sector Productivity and Cost Measures,” February 2008, http://www.bls.gov/lpc/ lpcmethods.pdf. 5. As a result of interest from data users, the BLS produces quarterly labor productivity measures and annual multifactor productivity measures for the total economy, but they are known to be downward biased and are therefore not included in the press releases. See “A Guide to the National Income and Product Accounts of the United States,” http://www.bea. gov/national/pdf/nipaguid.pdf for a description of output measures. 6. Note that measured labor productivity growth can still be positive if, for example, the composition of the labor force shifts toward more skilled (i.e., higher-paid) workers or if these sectors become more capital intensive. 7. To facilitate comparisons across various time periods, quarterly estimates are expressed as annualized levels, and quarterly growth rates are expressed as annual growth rates using the X

4

following formula: g t = t − 1 . Xt −1 8. These include depository institutions, nondepository institutions, security and commodity brokers, insurance carriers, regulated investment offices, small business offices, and real estate investment trusts. 9. The press release includes quarterly and annual indexes, and percentage changes, for output per hour and related measures, such as unit labor costs, real and current dollar compensation per hour, and unit nonlabor payments. See http://www.bls.gov/lpc/home.htm.

146 Lucy P. Eldridge, Chris Sparks, and Jay Stewart 10. BLS has looked into publishing quarterly estimates of labor productivity by industry, but the series is fairly volatile and BLS is not publishing these data at this time. See Eldridge (2016). 11. Industry measures produced include levels, annual indexes, and percentage changes for output per hour, output per employee, output, implicit price deflators for output, employment, hours of employees, labor compensation, and unit labor costs. Separate news releases are issued for selected services (http://www.bls.gov/news.release/prin2.nr0.htm), manufacturing (http://www.bls.gov/news.release/prin.nr0.htm), and trade (http://www. bls.gov/news.release/prin1.nr0.htm). 12. In order to have a complete industry accounting of the private business sector, estimates of MFP for nonmanufacturing industries corresponding to the NIPA level are prepared with available data. For several industries, there are concerns about the quality of the measures, and therefore the nonmanufacturing measures are not included in official BLS news releases. 13. Internationally, value-added labor productivity is by far the most frequently computed productivity statistic. See OECD (2001). This measure is easy to calculate and accounts for all factors of production except capital and labor composition. 14. For more discussion, see Gullickson (1995). 15. For more information on how imports affect productivity measures, see Eldridge and Harper (2010). 16. The BLS is conducting research on the feasibility of constructing new productivity measures using alternative output concepts. 17. For a complete discussion to the advantages and disadvantages of the gross and value- added output concepts, see OECD (2001), Chapter 3, 23–33. 18. Note that for four- digit NAICS industries, it is not possible to account for labor composition. 19. See Dean, Harper, and Sherwood (1996). 20. If technical change within an industry does not affect all factors of production but operates primarily on the primary inputs, then value-added approach is preferable (OECD 2001, 28). 21. Thus, labor productivity measures using value-added are less sensitive to outsourcing than sectoral output labor productivity, although the reverse is true when measuring multifactor productivity. Multifactor productivity measured using a sectoral output concept is less sensitive to outsourcing than value-added based measures. 22. See Gullickson (1995). 23. See Chapter 11 of BLS Handbook of Methods for more information on industry measures. 2 4. Output and the corresponding inputs for these particular nonmanufacturing industries are difficult to measure and can produce productivity measures of inconsistent quality. Multifactor productivity is estimated for these industries because of interest from data users; however, data users are alerted to be cautious when interpreting these data. 25. This is also known as an RAS procedure. See Denton (1971). The Denton proportional first difference method preserves the pattern of growth in quarterly indicator series by minimizing the proportional period-to-period change while meeting the average annual level constraints. 26. Labor composition indexes are only calculated for major sector and NIPA-industry MFP data—labor composition indexes are not calculated for four-digit NIACS industries due to source data limitations.

The US Bureau of Labor Statistics Productivity Program 147 27. Hours worked also include paid time for traveling between job sites, coffee breaks, and machine downtime. They also include time spent on maintenance activities. 28. Many countries use household data to estimate labor productivity, mainly because establishment data are not available. Some studies have examined how labor productivity growth would differ if CPS data were used exclusively. See Ramey (2012), Cociuba et al. (2012), and Burda et al. (2013), who also look at ATUS data. 29. For more information on the CES, see http://www.bls.gov/ces/. 30. For more information on the CPS, see http://www.bls.gov/cps/. 31. For more information on the NCS, see http://www.bls.gov/ncs/. 32. The CES all-employee hours series begins in March 2006, and BLS is currently evaluating whether to start using this new series. 33. In goods-producing industries, workers are divided into production and nonproduction workers. Nonproduction workers include professional, specialty, and technical workers; executive, administrative, and managerial workers; sales workers, and administrative support workers, including clerical. In service-providing industries, workers are divided into supervisory and nonsupervisory workers. Supervisory workers include all executives and managerial workers. 34. The Census Bureau’s industry codes, which are used in the CPS, have become more consistent with the NAICS codes used in the establishment survey. 35. The BLS major-sector productivity program uses annual hours-worked to hours-paid ratios at the three-digit industry level in manufacturing and more aggregate data outside of manufacturing. The quarterly series is estimated using a Denton procedure. Industry level data below the three-digit level are constructed using the three-digit ratios wherever possible. 36. This methodology for estimating nonproduction/supervisory hours was introduced in August 2004. See Eldridge, Manser, and Otto (2004). 37. For major sectors, hours worked by employees are calculated at the three-digit NAICS industry level; for detailed industries, employee data are constructed at a more detailed level. Seasonally adjusted monthly data from the CES are used to construct quarterly averages of employment and average weekly hours. National Compensation Survey data are used at the three-digit level where available. 38. The published quarterly labor productivity statistics aggregate employee hours from 14 major industry groups and then adds aggregate hours worked by self-employed and unpaid family worker hours. For quarterly measures, this ratio is constructed by pooling three months of data, seasonally adjusted using an X-12-ARIMA program. The BLS performs indirect seasonal adjustment (seasonally adjusting the components of the hours calculation rather than the final value), which is preferred when component series are suspected of having distinct seasonal patterns. Given the limited observations for some industry groups, the CPS data are seasonally adjusted quarterly rather than monthly. 39. See Stewart (2014) for a longer discussion of the issues and a summary of the relevant research. 40. This should not be too surprising, because the reference week (the week that includes the 12th of the month) was chosen to avoid holidays. 41. The BLS only constructs labor composition estimates for the multifactor productivity measures for the private business sectors, manufacturing sector, three- digit manufacturing industries, and NIPA nonmanufacturing industries.

148 Lucy P. Eldridge, Chris Sparks, and Jay Stewart 42. Households in the CPS are surveyed for four consecutive months, are out of the survey for eight months, and then resurveyed for four additional months. Earnings data are collected only in months-in-sample 4 and 8, which are referred to as the outgoing rotations. 43. The BLS recently switched from using the March Annual Social and Economic (ASEC) Supplement to using the ORG data. The main advantages of the ORG data are (1) a much larger sample size (two to three times the size of the ASEC sample depending on the year), and (2) better concordance between earnings and workers’ industry classification. The main disadvantages are (1) there are no earnings data on second jobs, and (2) there are no earnings data on self-employed workers. The productivity program assumes that self-employed workers and workers on second jobs have the same hourly wage as demographically similar wage and salary workers on their main jobs. Although the ASEC data are more comprehensive in that they include data on earnings and hours from all jobs, including self-employment, the ASEC only collects industry data for the longest job held in the previous year, which means that some earnings and hours are not attributed to the correct industry when the individual held more than one job during the year. To further increase sample size, BLS will start using the entire CPS (that is, adding months in sample 1–3 and 5–7) to estimate changes in hours worked. 44. Additional information concerning data sources and methods of measuring labor composition can be found in Zoghi (2010). 45. Data on investments in physical assets are obtained from BEA. See Bureau of Labor Statistics [2009], Multifactor Productivity Trends, 2007, News Release, US Department of Labor, #09-0302 (March 25). 46. Land and inventories capital inputs are also calculated using productive stocks and estimated rental prices. However, land and inventories stock are estimated directly, rather than via the perpetual inventory method, and rental prices do not have a depreciation term. 47. BEA and many academic studies assume geometric deterioration, where the rate of deterioration is constant over time. This simplifies calculations, because this deterioration rate can be applied to the entire stock of capital, so there is no need to keep track of assets’ vintages. However, Harper (1982, 1999) concluded that this assumption was unrealistic in many cases because assets’ productivity tend to decline slowly at first and more rapidly as they approach the end of their service lives. 48. A one-hoss shay asset functions the same until the end of its service life, at which time it fails (from the Oliver Wendell Holmes, Sr. poem “The Deacon’s Masterpiece or the Wonderful One-Hoss Shay”). 49. These values were chosen because they are close to values estimated by Hulten and Wykoff (1981a) using actual data. 50. The BEA provides the BLS with information on service lives and depreciation rates. The BLS does not use the BEA service lives. Instead, the BLS estimates services lives so that, when used in conjunction with the hyperbolic age-efficiency function, they are consistent with the BEA’s depreciation rates. 51. Because average maximum service lives are estimates and there are no data about individual asset types within categories, the distribution can also be viewed as accounting for the uncertainty, as well as heterogeneity. 52. Depreciation is calculated as the change in wealth stock minus the current year’s investment. The depreciation rate is equal to depreciation divided by the wealth stock, although in practice the BLS uses the productive capital stock because it is less volatile than the wealth stock. The present value of $1 of tax depreciation allowance is calculated using equations from Hall and Jorgenson (1967), and the deflator is calculated from the real and nominal investment data the BLS gets from the BEA. Use of the internal rate of return is

The US Bureau of Labor Statistics Productivity Program 149 necessary because of the assumption of perfect competition, which requires that profits be zero in equilibrium. 53. Note that equation (3.16) is defined for an industry × asset combination, whereas the rental price in equation (3.19) is defined for an industry only. Since capital income is available only at the industry level, equation (3.16) is modified using equation (3.21) so that the rental price used in the derivation of equation (3.20) is defined at the industry level. 54. This description glosses over many details. The Törnqvist aggregation is done in two steps, with the BEA asset categories being aggregated into seven asset groups for each industry (equipment, structures, land, rental residential capital, inventories, intellectual property products, and “all assets”). These asset group × industry combinations are then combined using the Törnqvist formula to arrive at the required intermediate and final aggregates. 55. Capital input can also be decomposed by type of asset, as in Table 3.1, which shows the portion of capital input attributable to each type of asset. 56. The data are adjusted from a cash basis to an accrual basis, so that the reference period is consistent with the output data. 57. See Giandrea and Sprague (2017) for a description of how the BLS measures labor’s share. 58. For more on the compensation gap, see Fleck, Glaser, and Sprague (2011). 59. The BLS cannot use the BEA depreciation and service life data directly, because the BLS and BEA assume different age-efficiency functions. The BEA assumes a constant (geometric) rate of depreciation/deterioration, whereas the BLS’s hyperbolic function assumes that the rate of deterioration increases as the asset ages. To account for this difference, the BLS recalculates average maximum asset service lives (Ω) so that the average deterioration rate is close to the BEA’s constant rate. The resulting BLS estimates of service lives tend to be a little longer than those of the BEA. 60. BEA obtains service lives from these sources, and calculates depreciation rates (Fraumeni 1997; Giandrea 2015). 61. This line of research uses data from the ASM or data from other countries (for example, see Andrews, 2015). The US data reside at the Census Bureau, and access is restricted (this restriction also applies to BLS employees)—they can be used only for approved projects. This is major drawback of the decentralized US statistical system. It is very difficult for the different statistical agencies to share data. 62. For more discussion, see Gullickson (1995). 63. MFP measures are available for all NIPA-level services-providing industries in response to data users’ demand. However, data users are cautioned that output and the corresponding inputs for nonmanufacturing industries are often difficult to measure and can produce productivity measures of inconsistent quality. Data users should be cautious when interpreting these data. 64. The BEA output measure for this industry is generated from data on labor compensation and other inputs. 65. Aaronson and Figura (2010) find evidence of this.

References Aaronson, Stephanie, and Andrew Figura. 2010. “How Biased are Measures of Cyclical Movements in Productivity and Hours?” Review of Income and Wealth 56(3): 539–558. Abraham, Katharine G., James R. Spletzer, and Jay C. Stewart. 1998. “Divergent Trends in Alternative Wage Series.” In Labor Statistics Measurement Issues, edited by John Haltiwanger, Marilyn Manser, and Robert Topel, 293–325. Chicago: University of Chicago Press.

150 Lucy P. Eldridge, Chris Sparks, and Jay Stewart Abraham, Katharine G., James R. Spletzer, and Jay C. Stewart. 1999. “Why Do Different Wage Series Tell Different Stories?” American Economic Review Papers and Proceedings 89(2): 34–39. Andrews, Dan, Chiara Criscuolo, and Peter N. Gal. 2015. The Global Productivity Slowdown, Technology Divergence and Public Policy: A Firm Level Perspective. Paris: OECD. Burda, Michael C., Daniel S. Hamermesh, and Jay Stewart. 2013. “Cyclical Variation in Labor Hours and Productivity Using the ATUS.” American Economic Review Papers & Proceedings 103(3): 99–104. Champagne, Julien, Andre Kurmann, and Jay Stewart. 2015. “Reconciling the Divergence in Aggregate U.S. Wage Series.” Labour Economics 49, December: 27–41. Cociuba, Simona E., Edward C. Prescott, and Alexander Ueberfeldt. 2012. “U.S. Hours and Productivity Using CPS Hours Worked Data: 1947-III to 2011-IV.” Unpublished manuscript. Cunningham, Cindy, Lucia Foster, Cheryl Grim, John Haltiwanger, Sabrina Wulff Pabilonia, Jay Stewart, Zoltan Wolf. 2018. “Dispersion in Dispersion: Measuring Establishment-Level Differences in Productivity.” Working Paper 18–25, Center for Economic Studies, US Census Bureau. Dean, Edwin R., Michael J Harper, and Mark S Sherwood. 1996. “Productivity Measurement with Changing Weight Indices of Outputs and Inputs.” In Industry Productivity: International Comparisons and Measurement Issues, edited by Bart van Ark, 183–215. Paris: OECD. Dean, Edwin R., and Michael J Harper. 2001. “The BLS Productivity Measurement Program.” In New Developments in Productivity Analysis, edited by Charles R. Hulten, Edwin R. Dean, and Michael J. Harper, 55–84. Chicago: University of Chicago Press. Denton, Frank T. 1971. “Adjustment of Monthly or Quarterly Series to Annual Totals: An Approach Based on Quadratic Minimization.” Journal of the American Statistical Association 66 (333): 99–102. Eldridge, Lucy P. 2016. “Measuring Quarterly Labor Productivity by Industry.” Monthly Labor Review. http://www.bls.gov/opub/mlr/2016/article/measuring-quarterly-labor- productivity-by-industry.htm. Eldridge, Lucy P., and Michael J. Harper. 2010. “Effects of Imported Intermediate Inputs on Productivity.” Monthly Labor Review 133(6): 3–15. Eldridge, Lucy P., Marilyn E. Manser, and Phyllis F. Otto. 2004. “Alternative Measures of Supervisory Employee Hours and Productivity Growth.” Monthly Labor Review 127(4): 9–28. Eldridge, Lucy P., and Sabrina Pabilonia. 2010. “Bringing Work Home: Implications for BLS Productivity Measures.” Monthly Labor Review 133(12): 18–35. Elsby, Michael W. L., Bart Hobijn, and Aysegul Şahin. 2013. “The Decline of the U.S. Labor Share.” Brookings Papers on Economic Activity 44(2): 1–63. Fleck, Susan, John Glaser, and Shawn Sprague. 2011. “The Compensation-Productivity Gap: A Visual Essay. Monthly Labor Review 134(1): 57–69. Fraumeni, Barbara. 1997. “The Measurement of Depreciation in the U.S. National Income and Product Accounts.” Survey of Current Business (July): 7–23. Frazis, Harley, and Jay Stewart. 2004. “What Can Time Use Data Tell Us about Hours of Work?” Monthly Labor Review 127(12): 3–9. Frazis, Harley, and Jay Stewart. 2009. “Comparing Hours Worked per Job in the Current Population Survey and the American Time Use Survey.” Social Indicators Research 93(1): 191–195. Frazis, Harley, and Jay Stewart. 2010. “Why Do BLS Hours Series Tell Different Stories about Trends in Hours Worked?” In Labor in the New Economy, edited by Katharine G. Abraham,

The US Bureau of Labor Statistics Productivity Program 151 James R. Spletzer, and Michael J. Harper, 343–372. NBER Studies in Income and Wealth. Chicago: University of Chicago Press. Giandrea, Michael, and Shawn Sprague. 2017. “Estimating Labor Share.” Monthly Labor Review, February. https://www.bls.gov/opub/mlr/2017/article/estimating-the-us-labor-share.htm Giandrea, Michael. 2015. “Estimation of Asset Service Lives by the Bureau of Labor Statistics.” Internal BLS report. Gullickson, William. 1995. “Measurement of Productivity Growth in U.S. Manufacturing.” Monthly Labor Review 118(7): 13–28. Hall, R. E., and Dale W. Jorgenson. 1967. “Tax Policy and Investment Behavior.” American Economic Review 57(3): 391–414. Harper, Michael J. 1982. “The Measurement of Productive Capital Stock, Capital Wealth, and Capital Services.” BLS Working Paper 128, June. Harper, Michael J. 1999. “Estimating Capital Inputs for Productivity Measurement: An Overview of U.S. Concepts and Methods.” International Statistical Review 67(3): 327–337. Hulten, Charles R., and Frank C. Wykoff. 1981a. “The Estimation of Economic Depreciation Using Vintage Asset Prices.” Journal of Econometrics 15: 367–396. Hulten, Charles R., and Frank C. Wykoff. 1981b. “The Measurement of Economic Depreciation.” In Depreciation, Inflation, and the Taxation of Income from Capital, edited by Charles R. Hulten, 81–125. Urban Institute Press, Washington, DC. OECD. 2001. Measuring Productivity: Measurement of Aggregate and Industry-Level Productivity Growth. Paris: OECD. Powers, Susan G. 2016a. “Labor Productivity Growth in Elementary and Secondary School Services: 1989–2012.” Monthly Labor Review, June. http://w ww.bls.gov/opub/ mlr/2 016/article/l abor-productivity-g rowth-i n-e lementary-and-s econdary-s chool- services.htm. Powers, Susan G. 2016b. “Heterogeneous Education Outputs in Public School Elementary and Secondary Education Output Measures: Disabled and Non-Disabled Students.” Monthly Labor Review, September. https://www.bls.gov/opub/mlr/2016/article/heterogeneous- education-output-measures-for-public-school-students-with-and-without-disabilities. htm Ramey, Valerie. 2012. “The Impact of Hours Measures on the Trend and Cycle Behavior of U.S. Labor Productivity.” Unpublished paper, University of California, San Diego. Royster, Sara E. 2012. “Improved Measures of Commercial Banking Output and Productivity.” Monthly Labor Review 135(7): 3–17. Solow, Robert M. 1957. “Technical Change and the Aggregate Production Function.” Review of Economics and Statistics 39(3): 312–320. Stewart, Jay. 2014. “The Importance and Challenges of Measuring Work Hours.” IZA World of Labor. http://wol.iza.org/articles/importance-and-challenges-of-measuring-work-hours/ long. Sveikauskas, Leo, Samuel Rowe, James Mildenberger, Jennifer Price, and Arthur Young. 2015. “Productivity Growth in Construction.” BLS Working Paper 478. Sveikauskas, Leo, Samuel Rowe, James Mildenberger, Jennifer Price, and Arthur Young. 2016. “Productivity Growth in Construction.” Journal of Construction Engineering and Management. 142(10). http://ascelibrary.org/doi/abs/10.1061/%28ASCE%29CO.1943-7862.0001138. Sveikauskas, Leo, Samuel Rowe, James Mildenberger, Jennifer Price, and Arthur Young. 2018. “Measuring Productivity Growth in Construction.” Monthly Labor Review. https://www.bls. gov/opub/mlr/2018/article/measuring-productivity-growth-in-construction.htm

152 Lucy P. Eldridge, Chris Sparks, and Jay Stewart Syverson, Chad. 2011. “What Determines Productivity?” Journal of Economic Literature 49(2): 326–365. US Bureau of Labor Statistics. “Chapter 11: Industry Productivity Measures,” BLS Handbook of Methods. https://www.bls.gov/opub/hom/pdf/homch11.pdf Zoghi, Cynthia. 2010. “Measuring Labor Composition: A Comparison of Alternate Methodologies.” In Labor in the New Economy, edited by Katharine G. Abraham, James R. Spletzer, and Michael J. Harper, 457– 485. NBER Studies in Income and Wealth. Chicago: University of Chicago Press.

Chapter 4

Theoret i c a l Produ ctivit y I ndi c e s R. Robert Russell

4.1. Introduction: Solow Technical Change This chapter focuses on theoretical productivity indices, defined for the purpose at hand as indices predicated on the assumption that the technology is known and nonstochastic but unspecified (e.g., nonparameterized).1 See Chapter 2 of this Handbook for a thorough discussion of more expansive notions of nonstochastic productivity indices. I begin, in this introductory section, with a preview of the chapter, motivating the contents with a retrospective look at Solow’s (1957) seminal contribution to the measurement of technical change, the genesis of today’s analysis of productivity comparisons (across production units as well as time). A (mostly inconsequential) generalization of Solow’s setup2 is as follows:

y t = f t (x t ) ∀ t ,

where, for observation t, y t ∈R + is the scalar output quantity, x t ∈R n+ is an input quantity vector, and f t : R n+ → R + is the production function. The nomenclature of the chapter adheres to Solow’s context in which t is a time index, but the results apply equally to cross-sectional analysis, where t is an index assigned to different economic units (e.g., firms or national economies). For expositional purposes only, Solow considered the special case of neutral technological change:

y t = at f (x t ) ∀ t , (4.1)

154 R. Robert Russell where at is a scalar indicator of the state of technology and f : R n+ → R + is a stationary (reference) production function. Additionally, at = y t / f (x t ) has come to be interpreted as total factor productivity at observation t, where f (x t ) is the (aggregate) total factor input at observation t. Solow’s definition of technical change from a base period b to a current period c, under the assumption of neutral technological change, is given simply by the relative productivity states, ac / ab . The Solow productivity index under neutrality, Π SN : R 2+(n +1) → R + , is then obtained by substitution from (4.1):3

Π SN (x b , x c , y b , y c ) :=

ac y c / f (x c ) . (4.2) = ab y b / f ( x b )

The last term in (4.2) underscores the interpretation of Solow’s technical change index, under the assumption of neutral technological change, as an index of total-factor productivity change. Another evocative interpretation of Solow technical change is obtained by multiplying top and bottom of ac / ab by f (x b ) to obtain

Π SN (x b , x c , y b , y c ) :=

ac f (x b ) f c (x b ) = . (4.3) ab f ( x b ) yb

The last term indicates that Solow technical change is given by the ratio of (a) the maximal (frontier) output that can be produced with the base-period input vector using the current-period technology to (b) the (frontier) output produced in the base period. While the last equality in (4.3) holds only under the assumption of neutral technology change, the last term is a natural definition of technological change for the general case of (possibly) non-neutral technological change. Under neutrality, the proportional shift of the frontier is identical across input vectors; under non-neutrality, when the frontier shift is input dependent, the last term in (4.3) measures the shift in the frontier specifically at the base-period input vector. In fact, as Solow did not assume neutrality in his calculations, his index was actually defined by

Π Sb (x b , x c , y b , y c ) :=

f c (x b ) , (4.4) yb

which is input-quantity dependent and indicates how much more can be produced with the base-period input vector using the current-period technology (rather than the base- period technology). While Solow measured technological change at the base-period input quantity (using the current-period technology), he could just as well have measured it at the current- period input vector (using the base-period technology):

Π Sc (x b , x c , y b , y c ) :=

yc . (4.5) f b (x c )

Theoretical Productivity Indices 155 The indices in (4.4) and (4.5) are identical if and only if technological change is neutral; that is, if and only if the production function is given by (4.1). The indices (4.4) and (4.5) are special cases—owing to the restriction to a single output—of the widely employed (output-oriented) Malmquist productivity indices formulated by Caves, Christensen, and Diewert (1982) (CCD). Generalization of (4.4) and (4.5) to multiple outputs requires an aggregation rule to obtain a measure of the change in “output.” The aggregation rule employed by CCD is the Malmquist (1953) output distance (gauge) function (independently formulated by Shephard 1953). In this context, it serves essentially as a measure of the “radial distance” from an output vector to a future or past technology frontier in output space (i.e., a production possibility frontier) for a fixed input, much as, for example, y c /f b (x c ) is a proportional measure of the distance from current-period output to the base-period production frontier at the current-period input vector x c. Measuring technological change in the region of 〈input, output〉 space at which the economic unit is operating has intuitive appeal, but there seems to be no compelling criterion to choose between using the current-period technology and the base-period quantities (4.4) or using the base-period technology and the current-period quantities (4.5). Diewert (1992a) therefore suggests adoption of the Fisher “ideal” index number formulation—namely, taking the geometric average of the two indices. In the single- output case, the Malmquist “ideal” output-based productivity index is

f c (x b ) Π (x , x , y , y ) = y b MI

b

c

b

c

1/ 2

yc f b (x c )

1/ 2

.

As the preceding nomenclature suggests, Diewert also formulated input-based Malmquist and Malmquist “ideal” indices. These formulations employ the Malmquist- Shephard input-oriented distance function, a radial measure of the distance of an input vector from a base-period or current-period frontier in input space (i.e., an isoquant). The CCD-Diewert development of the various Malmquist productivity indices in the general case of multiple outputs and multiple inputs is described in section 4.3 (following presentation of some preliminaries in section 4.2). Suppose we multiply top and bottom of the indices in (4.4) and (4.5) by y c = f c (x c ) and y b = f b (x b ), respectively, to obtain

Π Sb (x b , x c , y b , y c ) =

yc / yb (4.6) f c (x c ) / f c (x b )

Π Sc ( y c , y b , x c , x b ) =

yc / yb . (4.7) f (x b ) / f b (x c )

and

b

156 R. Robert Russell Thus, the Solow index can also be interpreted as a measure of the change in output divided by a measure of the change in an aggregate input, where the input aggregation rule is the current- period production function in (4.6) and the base-period production function in (4.7).4 Following up on a suggestion of Diewert (1992a), Bjurek (1996) exploited the CCD theory of Malmquist indices of aggregate input and output change (based respectively on Malmquist-Shephard input and output distance functions) to develop multiple- output (as well as multiple-input) productivity indices analogous to (4.6) and (4.7).5 Diewert attributed the ideas behind this index formulation to Moorsteen (1961) and (maybe) Hicks (1961), and the Hicks-Moorsteen productivity index assignation has stuck. These indices are exposited in section 4.4. The CCD and Bjurek papers are essential to understanding most of today’s research on multiple-output productivity, and the remainder of this chapter covers extensions and refinements of the Malmquist and Hicks-Moorsteen indices. The possibility of technological inefficiency—production below the frontier—is implicit in the CCD-Bjurek framework, since the Malmquist-Shephard input and output distance functions serve as measures of inefficiency as well as shifts in the production frontier. Exploiting this dual use of the distance functions, Färe, Grosskopf, Norris, and Zhang (1994) decompose the Malmquist index into two components: technological change (shifts in the production frontier) and efficiency change (movements toward or away from the frontier). This decomposition is discussed in section 4.5. Much—perhaps most—of the theoretical research on productivity indices in recent years has been directed at the incorporation of additional decomposition components, like scale effects and input-or output-mix effects; this literature is also reviewed in section 4.5. The Malmquist and Hicks-Moorsteen indices, owing to their radial structure, are not applicable to the measurement of productivity in the full space of inputs and outputs.6 Section 4.6 describes two nonradial indices: the hyperbolic index and the (Luenberger 1992) directional-distance index. Dual productivity indices (employing cost and revenue functions) and aggregation of productivity indices across economic units are discussed in sections 4.7 and 4.8, respectively. In the concluding comments in section 4.9, I take liberties—for the most part resisted in earlier sections—to offer some evaluative comments, particularly with respect to the comparison of the Malmquist and Hicks-Moorsteen approaches.

4.2. Preliminaries: Technological Change and Productivity Indices 4.2.1. Technologies Although I frequently refer to the unit of analysis as a “firm,” the theory is applicable to any type of economic unit, even an aggregate (national) economy. The firm produces a

Theoretical Productivity Indices 157 vector of outputs y ∈R m+ using a vector of inputs x ∈R n+ . To avoid the nuisance of dealing with null vectors, some of our functions map from production space with the origins of input and output space expunged:7 R n++ m = R n+ \ {0[n] } × R m+ \ {0[m] } =: R n+ × R m+ . The firm’s production is constrained by a (known) technology—the set of technologically feasible production vectors,

T = {〈 x , y 〉 ∈R n++ m | x can produce y}.

Given the technology T, the output-possibility set for input vector x is8

P (x , T ) = { y ∈R m+ | 〈 x , y 〉 ∈T }

and the input-requirement set for output vector y is

L( y , T ) = {x ∈R n+ | 〈 x , y 〉 ∈T }.

Clearly, 〈 x , y 〉 ∈T ⇔ x ∈L( y , T ) ⇔ y ∈P (x , T ). We restrict the set of technologies, denoted , to those that are closed and satisfy free input disposability (L( y , T ) + R n+ = L( y , T ) for all y ∈R m+ ) and free output disposability (P (x , T ) = (P (x , T ) − R m+ ) ∩ R m+ for all x ∈R n+ ). As T is closed for all T ∈ , so are the slices, L( y , T ) and P (x , T ), for all 〈 x , y 〉 ∈R n++ m. The isoquant for output y ∈R m+ is given by

I ( y , T ) = {x ∈L( y , T )| λ x ∉ L( y ) ∀ λ < 1},

and the production possibility surface for input x ∈R n+ is

Γ(x , T ) = { y ∈P (x , T )| λ y ∉ P (x ) ∀ λ > 1}.

4.2.2. Distance Functions An essential building block of multiple-output (and, of course, multiple-input) productivity analysis is the “distance function,” independently introduced into the economics literature,9 in different contexts and with different objectives, by Debreu (1951), Malmquist (1953), and Shephard (1953). Malmquist’s context is the closest to the thrust of this chapter (even though his objective was the construction of consumer cost-of- living indices rather than productivity indices). The input distance function, DI : N × → R + +, maps from a subset of production space, N = {〈 x , y 〉 ∈R n++ m | L( y , T ) ≠ ∅}, and the set of allowable technologies into the positive real line and is defined by

DI (x , y , T ) = max{λ > 0 | x / λ ∈L( y , T )}.

158 R. Robert Russell Thus, the input distance function is defined as the maximal (proportional) radial contraction (or minimal radial expansion) of a given input vector consistent with technological feasibility of production of a given output vector.10 Thus, the input distance function can be characterized as a notion of “radial distance” of an input-quantity vector from the frontier of the technology (the isoquant in input space)—a characterization that, as we shall see, is evocative in the construction of productivity indices. With our (parsimonious) restrictions on the technology (closedness and free input disposability), DI is well-defined, homogeneous of degree one and nondecreasing in x, and nonincreasing in y for all x , y , T ∈N × . Moreover, x ∈L( y , T ) ⇔ DI ( y , x , T ) ≥ 1, and x ∈ I ( y , T ) ⇔ DI (x , y , T ) = 1, so that, for any y ∈R m+ , L( y , T ) is recovered from DI by L( y , T ) = {x ∈R n+ | DI (x , y , T ) ≥ 1}

and I ( y , T ) is recovered from DI by I ( y , T ) = {x ∈R n+ | DI (x , y , T ) = 1}.

Thus, the input distance function also serves as a functional representation of the technology. The output distance function, DO : N × → R + , is similarly defined by

DO (x , y , T ) = min{λ > 0 | y / λ ∈P (x , T )};

that is, as the minimal (proportional) radial contraction (or maximal radial expansion) of a given output vector consistent with technological feasibility of production for a given input vector. Thus, the output distance function is the “radial distance” of an output-quantity vector from the frontier of the technology (the production possibility surface in output space).11 With our restrictions on the technology (closedness, free output disposability, and boundedness of P (x , T )), DO is well- defined, homogeneous of degree one and nondecreasing in y, and nonincreasing in x for all x , y ∈R n++ m. Moreover, y ∈P (x , T ) ⇔ DO (x , y , T ) ≤ 1, so that, for any x ∈R n+ , P (x , T ) is recovered by P (x , T ) = { y ∈R m+ | DO (x , y , T ) ≤ 1}

and Γ(x , T ) is recovered by

Γ(x , T ) = { y ∈R m+ | DO (x , y , T ) = 1},

indicating that the output distance function is a functional representation of the technology. Finally, it is easy to see that each of these distance functions is independent of units of measurement.12

Theoretical Productivity Indices 159

4.2.3. Technological Efficiency A firm is input inefficient at time t if DI (x t , y t , T t ) > 1 and output inefficient if DO (x t , y t , T t ) < 1. Moreover, with restriction of the distance- function domains to feasible production vectors 〈 x t , y t 〉 ∈T t , 1/ DI (x t , y t , T t ) =: EI (x t , y t , T t ) and DO (x t , y t , T t ) =: EO (x t , y t , T t ) are, respectively, (Debreu-Farrell) input and output efficiency indices (Debreu 1951; Farrell 1957). Each measures the radial distance of the quantity vector from the frontier and, for technologically feasible production vectors, maps into the (0,1] interval.

4.2.4. Technological Change The essence of technological comparisons across periods (or across production units) is the comparison of production possibilities, as reflected by production possibility sets or input requirement sets, under counterfactual assumptions about input availability or output requirement. The production possibility set at the current-period input vector using the base-period technology is P (x c , T b ), and the production possibility set at the base-period input vector using the current-period technology is P (x b , T c ). The corresponding production possibility surfaces are Γ(x c , T b ) and Γ(x b , T c ). If

P (x c , T b ) ⊂ P c (x c , T c ) ∧ Γ(x c , T b ) ∩ Γ(x c , T c ) = ∅, (4.8)

we say that, measuring qualitatively in output space and normalizing on the current- period input vector, the technology of the economic unit unambiguously (globally) improved between b and c. This normalization is arbitrary, and we could just as well normalize on the base-period input vector, in which case the technology of the economic unit unambiguously improved if

P (x b , T b ) ⊂ P (x b , T c ) ∧ Γ(x b , T b ) ∩ Γ(x b , T c ) = ∅. (4.9)

(These characterizations of unambiguous technological improvement can similarly be expressed in terms of [shrinking] input requirement sets at stipulated output vectors.) These criteria for global technological change are overly strong: a more reasonable requirement is to normalize on the output quantity vector as well—that is, to require only that the production possibility set expand in a neighborhood of the output quantity vector, y c in (4.8) or y b in (4.9). This is the approach adopted by CCD in their pathbreaking analysis of productivity measurement, entailing quantitative as well as qualitative measurement using the output distance function as a measure of the radial distance of an output vector in one period from the production possibility surface of the technology in the other period. The qualitative characterization of technological change in input space is analogous. The input requirement set at the current-period output vector using the base-period technology is L( y c , T b ), and the input requirement set at the base-period output vector

160 R. Robert Russell using the current-period technology is L( y b , T c ). The corresponding isoquants are I ( y c , T b ) and I ( y b , T c ). Under alternative normalizations, the technology of the economic unit unambiguously (globally) improves between b and c if L( y c , T c ) ⊂ L( y c , T b ) or L( y b , T c ) ⊂ L( y b , T b ). More reasonably, we can say that there has been technological progress (in the appropriate normalization) if the input requirement set expands (toward the origin) in a neighborhood of x c or, alternatively, x b .

4.2.5. Productivity Indices At the most abstract, generic (even austere) level, the definition of a multiple-output productivity index is a mapping, Π : R 2+(n + m) → R + + , with image Π x b , x c , y b , y c . Without some imposed structure or required set of properties, however, this concept is close to vacuous. The two main approaches to adding structure are the axiomatic, or test, approach (requiring that the index satisfy certain properties, like monotonicity) and the economic approach (tying the index to economic or technological constructs). (These approaches are complementary.) The desirable properties of productivity indices include the following:13

(

)

(I) Identity: 〈 x b , y b 〉 = 〈 x c , y c 〉 ⇒ Π(x b , x c , y b , y c ) = 1 . (M) Monotonicity: Productivity is nondecreasing in current-period output and base- period inputs and nonincreasing in current-period inputs and base-period outputs: 〈 x b , − x c , − y b , y c 〉 ≥ 〈 x b , − x c , − y b , y ) c

c b ⇒ Π(x b , x c , y b , y c ) ≥ Π(x , x c , y b , y ).

(UI) Invariance with respect to units of measurement (commensurability):

Π(x b K x , x c K x , y b K y , y c K y ) = Π(x b , x c , y b , y c ),

for arbitrary n × n and m × m positive diagonal (unit transformation) matrices, K x and K y . (P) Proportionality: Π is homogeneous of degree 1 in y c and homogeneous of degree –1 in x c (hence homogeneous of degree zero in 〈 x c , y c 〉 ). (T) Transitivity: For any three periods, b, c, and d, the product of the measured productivity changes from period b to period c and from period c to period d is equal to the productivity change from period b to period d :

Π(x b , x c , y b , y c ) ⋅ Π(x c , x d , y c , y d ) = Π(x b , x d , y b , y d ). 14

Theoretical Productivity Indices 161 Linkage to technological or economic concepts requires a more constructive approach. To this end, we expand the domain to encompass the space of technologies .15 The extended productivity index (in a slight abuse of notation), Π : R 2+(n + m) × 2 → R + +, now has the image Π(x b , x c , y b , y c , T b , T c ). In many specific cases, the index is invariant with respect to changes in some quantity vectors or one of the technologies. Application of the preceding axioms to indices with the enhanced domain is straightforward. (One could add axioms related to the technologies, but this approach, to my knowledge, has not been explored.)

4.3. Malmquist Productivity Indices Caves, Christensen, and Diewert (1982) (CCD) imposed structure by a natural linkage to the technology using the distance function as a representation of the technology and as a notion of distance from a quantity vector in one period to a technological frontier in another. They proposed four basic productivity indices predicated on alternative normalizations with respect to the choice of the base period or the current period for the reference technology and for the production vector. These indices are defined as follows (where the relation eff = holds only under the assumption of technological efficiency):16 Output-based, technology b-based Malmquist productivity index:

ΠOM (x b , x c , y b , y c , T b ) =

DO (x c , y c , T b )

DO (x , y , T b

b

b

)

eff

= DO (x c , y c , T b )

(4.10)

= min{λ > 0 | y c / λ ∈P (x c , T b )} =: ΠOM (x c , y c , T b ). Output-based, technology c-based Malmquist productivity index:

ΠOM (x b , x c , y b , y c , T c ) =

DO (x c , y c , T c ) eff 1 = b b c b DO (x , y , T ) DO (x , y b , T c )

( {

})

= min λ > 0 | y / λ ∈P (x , T ) b

b

c

−1

(4.11) M O

b

b

c

=: Π (x , y , T ).

Input-based, technology b-based Malmquist productivity index:

Π IM (x b , x c , y b , y c , T b ) =

DI (x b , y b , T b ) eff 1 = DI ( x c , y c , T b ) DI ( x c , y c , T b )

(

{

})

= max λ > 0 | x / λ ∈L( y , T ) c

c

b

−1

(4.12) =: Π (x , y , T ). M I

c

c

b

162 R. Robert Russell Input-based, technology c-based Malmquist productivity index:

DI (x b , y b , T c ) eff = D I (x b , y b , T c ) (4.13) D I (x c , y c , T c ) M b b c b b c = max{λ > 0 | x / λ ∈L( y , T )} =: Π I (x , y , T ).

Π IM (x b , x c , y b , y c , T c ) =

Thus, each of these productivity indices is defined, in the first instance, as a ratio of radial distances to an (input or output) frontier for base-period and current-period quantities and for a common technology.17 Under the assumption that production units operate (efficiently) on the production frontier, DO (x b , y b , T b ) = DI (x b , y b , T b ) = 1 for t = b, c , so that the indices simplify to single distance functions evaluated at a mixture of a production vector for one period and a technology for the other period. The (output-based) productivity index normalized on the base-period technology, defined in (4.10), measures the maximal radial contraction of the current-period output vector needed to place the contracted vector in the production possible set for current- period input, using the base-period technology. Of course, if y c ∈P (x c , T b ), it must be expanded radially to the frontier. Clearly, ΠOM (x c , y c , T b ) > 1 if and only if there is technological progress—expansion of the frontier—in the neighborhood of the current- period production vector; conversely, ΠOM (x c , y c , T b ) < 1 if and only if the frontier of the production possibility set has receded in this neighborhood, evincing technological retardation. Note that this index, under technological efficiency, is precisely a generalization to multiple outputs (and to nonhomothetic technologies) of Solow’s measure of technical progress in the scalar-output case (equation (4.5)).18 The calculation of the (output-based) productivity index normalized on the current- period technology in (4.11) generates the inverse of the minimal expansion of the base- period output vector required to reach the frontier of the current-period production possibility set for base-period input vector. ΠOM (x b , y b , T c ) > 1 if and only if there has been expansion of the frontier of production possibility set in the neighborhood of the base-period production vector and ΠOM (x b , y b , T c ) < 1 if and only if the frontier of the production possibility set has receded in this neighborhood. This index is a generalization of the scalar-output measure of technological change in (4.4).19 The (input-based) productivity index normalized on the base-period technology, defined in (4.12), measures the inverse of the maximal radial expansion of the current- period input vector needed to place the expanded vector in the input-requirement set for current-period output, using the base-period technology. Clearly, Π IM (x c , y c , T b ) > 1 if and only if there is technological progress—expansion of the isoquant toward the origin—in the neighborhood of the current-period production vector. This index is an alternative generalization of the scalar-output measure of technological change in (4.5).20 Interpretation of the (input-based) productivity index normalized on the current-period technology, defined in (4.13) is similar, and again Π IM (x b , y b , T c ) > 1 if and only if technological progress has lowered the isoquant in the neighborhood of base-period quantities.

Theoretical Productivity Indices 163 These four indices, in general, provide different implications about productivity change, even possibly about the direction of change. The delineations among them are arbitrary enough to suggest that there is unlikely to be a compelling criterion for choosing any one. One suggestion, inspired by Irving Fisher’s recommendation for dealing with arbitrary normalizations in the construction of index numbers, is to take geometric averages over current-period and base-period constructions. He referred to these confluences as “ideal” indices. In this spirit, Diewert (1992a) suggested “Fisher-ideal” indices for input-oriented and for output-oriented productivity indices; they are defined as follows: Malmquist “ideal” output-based productivity index: ΠOMI (x b , x c , y b , y c , T b , T c ) = (ΠOM (x b , x c , y b , y c , T b ) ⋅ ΠOM (x b , x c , y b , y c , T c ))

1/ 2

D (x c , y c , T b ) = O b b c DO (x , y , T )

(4.14)

1/ 2

eff

Malmquist “ideal” input-based productivity index: Π IMI (x b , x c , y b , y c , T b , T c ) = (Π IM (x b , x c , y b , y c , T b ) ⋅ Π IM (x b , x c , y b , y c , T c ))

1/ 2

D (x b , y b , T c ) = I c c b DI ( x , y , T )

eff

1/ 2

.

(4.15)

The properties of the input and output distance functions, described earlier, endow each of these indices with the identity (I), monotonicity (M), and unit-invariance (UI) properties. On the other hand, these indices satisfy neither proportionality (P) nor transitivity (T) on their domains.21 In the next section I describe an index that does satisfy proportionality (though not transitivity).

4.4. Hicks-Moorsteen Indices Recall from equation (4.2) that Solow technical change for a single output can be interpreted as a ratio of total factor productivity in two periods. The various Malmquist indices are indisputably measures of technological change—that is, shifts in the frontier of the technology—but in general they do not have the interpretation as ratios of aggregate productivity in the two periods. Nor can they be interpreted in general as ratios of (aggregate) output change to (aggregate) input change, unlike the Solow single-output index in the form given by identity (4.6) or (4.7). Building on the construction in CCD (and following a suggestion of Diewert 1992a), Bjurek (1996) developed multiple-output indices analogous to (4.6) and (4.7). The key

164 R. Robert Russell building blocks for these indices are essentially Malmquist distance functions evaluated at intertemporal mixtures of production vectors (base-period input vector and current- period output vector, or vice versa). Define the Malmquist quantity indices as follows (where, again, the last equation in each case holds only under the assumption that firms operate on the technological frontier): Technology b-based Malmquist output index:

QOM (x b , y b , x c , y c , T b ) =

DO (x b , y c , T b ) eff = DO (x b , y c , T b ). (4.16) DO (x b , y b , T b )

Technology c-based Malmquist output index:

M O

Q

(x , y b

b

c

c

, x , y ,T

c

)=

( )= 1 . (4.17) (x , y ,T ) D (x , y ,T )

DO x c , y c , T c DO

c

b

eff

c

O

c

b

c

Technology b-based Malmquist input index:

QIM (x b , y b , x c , y c , T b ) =

( ) = D x , y ,T . (4.18) ( ) D (x , y ,T ) DI x c , y b , T b I

b

b

b

eff

I

c

b

b

Technology c-based Malmquist input index:

QIM (x b , y b , x c , y c , T c ) =

DI (x c , y c , T c ) eff 1 = . (4.19) b c c b D I (x , y , T ) D I (x , y c , T c )

Identity (4.16) yields a measure of the (aggregate) output change between the base period b and the current period c. Specifically, it is the radial reduction of current-period output required to place the contracted output vector in the production possibility set for base-period input vector using the base-period technology. As such, it can be interpreted as the radial distance of the current-period output vector from the base- period production possibility set. Obviously, QO (x b , y b , x c , y c , T b ) > 1 if (and only if) y c ∉ P (x b , T b ) (i.e., the current-period output vector lies above the production possibility surface for base-period technology and input vector), in which case we say that output increased. Identity (4.17), on the other hand, yields the maximal expansion of the base-period output vector consistent with the expanded output vector remaining in the production possibility set for current-period input vector using the current-period technology. As such, it is a measure of the radial distance of the base-period output vector from the current-period technological frontier. Clearly, QO (x b , y b , x c , y c , T c ) > 1 if and only if y b is below the current-period output frontier, in which case we say that output increased. The input indices in (4.18) and (4.19) have analogous interpretations, in terms of radial distances of input vectors in one period to isoquants of the alternative period.

Theoretical Productivity Indices 165 The Hicks-Moorsteen productivity indices can now be defined as the ratios of output changes to input changes, normalized alternatively on base-period and current-period technologies:

Π HM (x b , x c , y b , y c , T b ) =

QOM (x b , x c , y b , y c , T b ) eff DO (x b , y c , T b ) = (4.20) QIM (x b , x c , y b , y c , T b ) DI (x c , y b , T b )

Π HM (x b , x c , y b , y c , T c ) =

QOM (x b , x c , y b , y c , T c ) eff DI (x b , y c , T c ) = . (4.21) QIM (x b , x c , y b , y c , T c ) DO (x c , y b , T c )

and

The Hicks-Moorsteen index can also be interpreted as a multiple-output generalization of the total-factor-productivity version of the Solow index, equation (4.2). Rewrite, for example, the first equality in (4.20) as Π HM (x b , x c , y b , y c , T b ) =

( ) ( ) . (4.22) (x , y ,T ) / D (x , y ,T )

DO x b , y c , T b / DI x c , y b , T b DO

b

b

b

I

b

b

b

The numerator is interpreted as the period-c ratio of a (static) output quantity index normalized on base-period input quantity and technology to an input quantity index normalized on base-period input quantity vector and technology. The denominator is similarly interpreted as an aggregate output–input ratio in the base period b. Finally, in the absence of a compelling criterion for choosing between the base-period and current-period normalizations, Bjurek suggests the “Fisher ideal” amalgamation to obtain the Hicks-Moorsteen “ideal” productivity index: Π HMI (x b , x c , y b , y c , T b , T c ) =

QOI (x b , x c , y b , y c , T b , T c ) QII (x b , x c , y b , y c , T b , T c )

, (4.23)

where

QOI (x b , x c , y b , y c , T b , T c ) = (QOM (x b , x c , y b , y c , T b ) ⋅ QOM (x b , x c , y b , y c , T c )) 1/ 2

and

QII (x b , x c , y b , y c , T b , T c ) = (QIM (x b , x c , y b , y c , T b ) ⋅ QIM (x b , x c , y b , y c , T c )) 1/ 2

are Fisher “ideal” quantity indices. A comparison of the Malmquist and Hicks- Moorsteen indices is instructive. In a nutshell, as first emphasized by Grifell-Tatjé and Lovell (1995), the former is a measure of technological change (shift in the production frontier), while the latter is a (broader) measure of the change in total-factor productivity (incorporating the effects of movement along the frontier as well as the shift of the frontier). Maintaining technological efficiency, each of the four Malmquist indices, (4.10)–(4.13), measures productivity

166 R. Robert Russell as the radial shift of the frontier (in either input or output space) in the neighborhood of a given production vector. The Hicks-Moorsteen indices, (4.20) and (4.21), on the other hand, incorporate the effects of changes in the production vector—most notably, scale effects22—as well as shifts in the technological frontier. This comparison suggests that the difference between the two types of indices vanishes if productivity is invariant with respect to shifts of the production vector along the surface of the technology. Indeed, this fact has been proved by Färe, Grosskopf, and Roos (1996). A straightforward generalization of their finding is as follows: if and only if, for all t, the technology satisfies (a) input and output homotheticity (isoquants and production possibility surfaces are radial transformations of one another)23 and (b) constant returns to scale (DO (λ x b , y b , T b ) = (1 / λ) ⋅ DO (x b , y b , T b ) for all 〈 x b , y b 〉 ∈T b), the four Malmquist indices, (4.10)–(4.13), and the two Hicks-Moorsteen indices, (4.20) and (4.21), are identical (cf. Balk 1998, 112–114).

4.5. Decompositions of Productivity Growth 4.5.1. Inefficiency and Technological Change Recall that CCD initially defined the four (output-and input-based, technology-based) Malmquist indices as ratios of distance functions. But under the assumption that production units operate (efficiently) on the production frontier, the indices simplify to eff the expressions to the right of the relation = in (4.10)–(4.13). Färe, Grosskopf, Norris, and Zhang (1994) (FGNZ) elaborated on the CCD framework by allowing firms to operate inefficiently, below the production possibility surface for the extant input vector or above the isoquant for the extant output vector. The acknowledgment that economic units might operate less than fully efficiently opens up the possibility of decomposing the Malmquist index into two components: changes in the technology (shifts of the frontier) and changes in efficiency (radial distance from the frontier). FGNZ executed this idea by rewriting the initial identity in (4.10)–(4.13) in terms of efficiency indices (see subsection 4.2.3) as follows:

ΠOM (x b , x c , y b , y c , T b ) =

ΠOM (x b , x c , y b , y c , T c ) =

DO (x c , y c , T b ) EO (x c , y c , T c )

DO (x c , y c , T c ) EO (x b , y b , T b )

, (4.24)

DO (x b , y b , T b ) EO (x c , y c , T c ) , (4.25) b b b DO (x b , y b , T c ) EO (x , y , T )

Theoretical Productivity Indices 167

Π IM (x b , x c , y b , y c , T b ) =

D I (x c , y c , T c ) E I ( x c , y c , T c ) , (4.26) DI ( x c , y c , T b ) E I ( x b , y b , T b )

Π IM (x b , x c , y b , y c , T c ) =

D I (x b , y b , T c ) E I ( x c , y c , T c ) . (4.27) b b b D I (x b , y b , T b ) E I ( x , y , T )

and

In each case, the Malmquist index allowing for the possibility of inefficient production is the multiple of two indices: the two ratios on the right-hand sides. The first ratio is an index of technical change: the radial shift, in either input space or output space, of the production frontier in the neighborhood of either the base-period or the current- period production vector. The second is an index of the change in efficiency, oriented alternatively to output space or input space. As in the case of efficient production, the output-oriented and input-oriented Malmquist “ideal” indices are defined as the geometric means taken over the indices normalized on the base period and the current period: ΠOMI (x b , x c , y b , y c , T b , T c )

(

)

(

= (ΠOM x b , x c , y b , y c , T b ⋅ ΠOM x b , x c , y b , y c , T c D (x c , y c , T b ) D ( x b , y b , T b ) = O ⋅ O c c c b b c DO (x , y , T ) DO (x , y , T )

1/ 2

))

1/ 2

EO (x c , y c , T c )

EO (x b , y b , T b )

and Π IMI (x b , x c , y b , y c , T b , T c )

= (Π IM (x b , x c , y b , y c , T b ) ⋅ Π IM (x b , x c , y b , y c , T c ))

1/ 2

D (x c , y c , T c ) D (x b , y b , T c ) = I c c b ⋅ I b b b D I ( x , y , T ) D I (x , y , T )

1/ 2

E I (x c , y c , T c ) . EI (x b , y b , T b )

Thus, as originally formulated, the CCD indices capture efficiency change as well as technological change, and FGNZ capitalized on this formulation to implement a binary decomposition of productivity change into these two components. The success of this decomposition, particularly in empirical studies, provided an impetus to refine the decomposition into additional contributing factors. These efforts are summarized in the next subsection.

4.5.2. Returns to Scale and Output Mix As noted earlier, the Malmquist productivity index does not encompass the effects of returns to scale (RTS) as quantities change. Thus, it is not surprising that efforts

168 R. Robert Russell to incorporate RTS into the decomposition of the Malmquist index24 led instead to revisions of the index itself. The first of these revisions is a “Malmquist index” defined, not on the actual technology, but instead on a virtual constant-returns-to-scale (CRS) technology, a conical envelopment of the actual technology. Among the many papers directed at this issue, perhaps the “constructive” approach—starting with individual components of a productivity index and then aggregating over these components to construct a consistent overall productivity index—best illustrates this progression. A good example is the contribution of Balk (2001), who referred to it as the “bottom-up” approach (as opposed to the “top-down” approach begun by FGNZ). Define the conical envelopment of the technology T by

(T ) = {〈 λ x , λ y 〉 ∈R n++ m | x , y ∈T ∧ λ > 0}

and restrict the set of allowable technologies to those for which (T ) is a proper subset of R n++ m (thus excluding, e.g., technologies with globally increasing returns to scale and Cobb-Douglas-like technologies where some marginal product goes to infinity as the quantity goes to zero). The cone (T ) is a virtual technology—informally, the “smallest” CRS technology containing T. Roughly speaking, scale efficiency at a point in production space is measured as the “proportional distance” from the actual technological frontier to the conical (envelopment) technological frontier in either input or output space. I stick here to the output notion; the alternative approach in input space can be found in Balk (2001). Output scale efficiency is thus defined as the ratio of output efficiency defined on the virtual technology divided by output efficiency defined on the true technology:

SO (x , y , T ) =

EO (x , y , (T )) DO (x , y , (T )) . = EO (x , y , T ) DO (x , y , T )

Note that SO (x , y , T ) ≤ 1 and, if the technology satisfies constant returns to scale, (T ) = T and S(x , y , T ) = 1. As in the case of the FGNZ decomposition of Malmquist productivity change into technical change and technological efficiency components, there exists more than one path of quantity and technological change along which to measure these components. Again in the face of space constraints, I consider only the path of, first, quantity changes along the base-period technological frontier and, then, shift of the frontier at current- level quantities. (Technological efficiency change is path independent.) The returns-to-scale index for a change in the input quantity vector from x b to x c is the resultant change in the gap between the actual and conical production frontiers:

SIO (x b , x c , y b , T b ) =

SO (x c , y b , T b ) DO (x c , y b , (T b )) / DO (x c , y b , T b ) . (4.28) = SO (x b , y b , T b ) DO (x b , y b , (T b )) / DO (x b , y b , T b )

Theoretical Productivity Indices 169 Interchanging the SW and NE part of the last term yields the interpretation of the index as a measure of the expansion (or contraction) of the virtual (conical) production possibility frontier along a ray through y b resulting from the change in the input vector divided by a measure of the expansion (or contraction) of the actual production possibility frontier along a ray through y b for the same change in the input vector. As Balk (2001) perspicaciously pointed out, (4.28) does not tell the whole story: as the input vector changes from x b to x c, the ray through the output vector also shifts to pass through y c (unless the technology satisfies output homotheticity). This shift also changes the level of scale efficiency, SO (x , y , T ).25 Balk refers to this change as the output- mix effect and measures it as follows:

MO ( x c , y c , y b , T b ) =

DO (x c , y c , (T b )) / DO (x c , y c , T b ) DO (x c , y b , (T b )) / DO (x c , y b , T b )

.

Combining the returns-to-scale and output-mix effects with the FGNZ technological and efficiency change indices in (4.24),

TCO (x c , y c , T b , T c ) :=

DO (x c , y c , T b ) DO (x c , y c , T c )

and

ECO (x b , x c , y b , y c , T b , T c ) :=

DO (x c , y c , T c ) , DO (x b , y b , T b )

we obtain, after some cancellation of terms, the overall productivity index,

TCO (x c , y c , T b , T c ) ⋅ ECO (x b , x c , y b , y c , T b , T c ) D (x c , y c , (T b )) ⋅SIO (x b , x c , y b , T b ) ⋅ MO (x c , y c , y b , T b ) = O b b . DO (x , y , (T b ))

Thus, the four components aggregate to a Malmquist index defined on the virtual (conical) technology (T ) rather than the actual technology T. The conical- envelopment approach to incorporating returns to scale into the Malmquist index has come under some criticism in recent years. I noted earlier the problem of the index not being defined for all technologies in . In addition, measuring technological change essentially by comparing “maximal average product” in two different periods, which has little if any economic significance, leaves much to be desired. An alternative bottoms-up approach is that of Peyrache (2014). Following up on suggestions by Lovell (2003), Peyrache constructed a discrete approximation to the differential returns-to-scale coefficient evaluated at y = y / DO (x , y , T ) (a radial projection of y to the output possibility surface):

170 R. Robert Russell (x , y ) =

∂ ln DO (λ x , y , T )−1 ∂ ln λ

. λ =1

Distinguishing between an approximation from below and from above (owing to possible nondifferentiability), Peyrache arrives at the following alternative indices measuring the contribution of returns to scale: c c DO (x b , y , T b ) / DO (x c , y , T b ) R(x , x , y , T ) = c c DI (x c , y , T b ) / DI (x b , y , T b ) b

c

c

b

b

b

and

b

c

R(x , x , y , T ) =

b

b

b

b

DO (x b , y , T b ) / DO (x c , y , T b ) DI (x c , y , T b ) / DI (x b , y , T b )

.

In each case, the denominator is an input quantity index normalized on the frontier b point y or y c . The numerator in each case is a measure of the radial expansion (or contraction if the value is less then 1) of the production possibility frontier along the ray through b y b—equivalently, through y (owing to the first-degree homogeneity of DO in y)—as the b input vector changes from x to x c.26 Thus, in each case, the returns-to-scale index is a discrete measure of the aggregate increase in frontier output along a ray through the projected period- t frontier point, y b or y c , as input quantities change between time b and time c. Peyrache then defines his radial productivity index, Π RPI , as the composition of the Malmquist output- oriented, technology- b - based productivity index (4.10)— which does not include scale effects—and the matching returns-to-scale index. Some manipulation, exploitating first-degree homogeneity of DO in y, yields the following: Π cRPI (x b , x c , y b , y c , T b ) = ΠOM (x b , x c , y b , y c , T b ) ⋅ R(x b , x c , y c , T b )

=

DO (x b , y c , T b ) / DO (x b , y b , T b ) c c DI (x c , y , T b ) / DI (x b , y , T b )

(4.29)

and ΠbRPI (x b , x c , y b , y c , T b ) = ΠOM (x c , x c , y b , y c , T b ) ⋅ R(x b , x c , y b , T b )

=

DO (x c , y c , T b ) / DO (x c , y b , T b ) . b b DI (x c , y , T b ) / DI (x b , y , T b )

(4.30)

The first index (4.29) normalizes on the current-period output ray, while the second (4.30) normalizes on the base-period output ray. As ΠOM (x c , x c , y b , y c , T b ) decomposes into a technological change component and an efficiency component, (4.29) and (4.30) yield tripartite decompositions of RPI productivity changes.

Theoretical Productivity Indices 171

4.6. Nonradial Indices and Indicators The orientation of the Malmquist productivity index to either input space or output space poses a quandary—which orientation to use when. An approach to circumvention of this quandary is the (singular) measurement of productivity in the full 〈input, output〉 space (commonly referred to as “graph space”). Measurement in this space cannot follow a radial path to a (current-or base-period) frontier, since outputs must be expanded and inputs must be contracted. The most natural extension of the Malmquist index to this space is the hyperbolic index, a multiplicative expansion of output and contraction of input to a (current-or base-period) frontier of graph space. An alternative is the directional distance indicator, an additive expansion of output and contraction of input in a stipulated direction to a current-or base-period frontier of graph space.27 These functions are discussed in the next two subsections.

4.6.1. Hyperbolic Indices As noted earlier, the distance-function components of Malmquist and Hicks-Moorsteen productivity indices are essentially equivalent to (simple transformations of) Debreu- Farrell technological efficiency indices. Extension of efficiency indices to 〈input, output〉 space is a precursor of extension of productivity indices to this space. This extension begins with the Färe, Grosskopf, and Lovell (1985) extension of the Debreu-Farrell efficiency index to 〈input, output 〉 space: E H (x , y , T ) = min{λ > 0 | 〈 λ x , y / λ 〉 ∈T }.

As the path of the production vector to the frontier (as λ shrinks to its minimal level) is hyperbolic, Färe, Grosskopf, and Lovell refer to this index as the hyperbolic efficiency index. It maps (technologically feasible) production vectors and technologies into the (0,1] interval and is nonincreasing in x and nondecreasing in y. Analogously to the use of the Debreu-Farrell efficiency index to formulate the Malmquist index in input or output space, the hyperbolic efficiency index can be used to construct a “hyperbolic” productivity index in 〈input, output 〉 space: Technology b-based hyperbolic Malmquist index:

Π

H

c

b

c

b

( ) = E x , y ,T . (4.31) ( ) (x , y ,T )

E H x c , y c ,T b

(x , x , y , y ,T ,T ) = E b

c

H

b

b

eff

H

c

c

b

b

Technology c-based hyperbolic Malmquist index:

(

)

ΠH xb , x c , yb , y c ,T b ,T c =

( )= 1 . (4.32) (x , y ,T ) E (x , y ,T )

E H x c , y c ,T c E

H

b

b

c

eff

H

b

b

c

172 R. Robert Russell Under the assumption of technologically efficient operation, this index measures productivity change as the “hyperbolic distance” from the current-period production vector to the base-period technology frontier or, alternatively, as the inverse of the hyperbolic distance from the base-period production vector to the current-period technological frontier. The index takes a value greater than one if and only if the technology frontier shifts upward in the hyperbolic direction in the neighborhood of the indicated production vector and hence is fundamentally a measure of technological change, analogously to the Malmquist index under the assumption of technological efficiency. These indices satisfy the identity (I), monotonicity (M), and unit-invariance (UI) properties, but fail to satisfy proportionality (P) or transitivity (T). The middle terms in (4.31) and (4.32) reflect the possibility of technological inefficiency: for example, E H (x b , y b , T b ) < 1 in (4.31) renders this term greater than the third term. Multiplication of top and bottom of these two terms by E H (x c , y c , T c ) and E H (x b , y b , T b ), respectively, yields decompositions of productivity change into indices of technical change and efficiency change:

Π H (x b , x c , y b , y c , T b , T c ) =

E H (x c , y c , T b ) E H (x c , y c , T c ) (4.33) E H (x c , y c , T c ) E H (x b , y b , T b )

Π H (x b , x c , y b , y c , T b , T c ) =

E H (x b , y b , T b ) E H (x c , y c , T c ) . (4.34) E H (x b , y b , T c ) E H (x b , y b , T b )

and

The first term on the right of each equation reflects technological change, measured in the hyperbolic direction and normalized alternatively on current-period and base- period production, while the second terms in each case reflect the change in efficiency. The hyperbolic Malmquist index was proposed and implemented by Zofio and Lovell (2001).

4.6.2. Luenberger (Directional Distance) Indicators The dominant nonradial formulation, however, has been the Luenberger productivity indicator, adapted from the shortage function of Luenberger (1992) to the measurement of productivity by Chambers (1996).28 In the economics literature, Luenberger’s concept has typically gone by the more evocative name, directional distance function (DDF). The DDF is defined by

D(x , y , T , g ) = max λ 〈 x − λ g x , y + λ g y 〉 ∈T ,

{

}

where g = 〈 g x , g y 〉 ∈R n++ m. This function measures the feasible 〈contraction, expansion 〉 of 〈input, output 〉 quantities in the direction g and maps into R + .

Theoretical Productivity Indices 173 Given our assumptions about the technology, the DDF is well-defined, nondecreasing in x, and nonincreasing in y for all x , y , T ∈N × . As g lies in the same coordinate space as the production vector, the DDF is also independent of units of measurement. Moreover, D(x , y , T , g ) ≥ 0 if and only if x , y ∈T , so that the DDF is a functional representation of the technology, and D(x , y , T , g ) = 0 if and only if x , y is contained in the frontier of T. In contrast to the Debreu-Farrell-Malmquist distance function, which measures distance in ratio form, the DDF measures distance in terms of vector differences (as multiples of the direction vector g ). Consequently, while the radial distance function has invariance properties with respect to certain rescaling of the data (owing to the homogeneity conditions), the DDF is invariant under transformations of the origin. The Luenberger productivity indicator is defined as the arithmetic average of differences of directional distances from production vectors in periods b and c to a common technology ( b or c): 1 b b c (D(x , y , T , g ) − D(x c , y c , T c , g ) 2 + D(x b , y b , T b , g ) − D(x c , y c , T b , g )) (4.35) 1 =: (Π cL (x b , x c , y b , y c , T c ) 2 + ΠbL (x b , x c , y b , y c , T b ));

Π L (x b , x c , y b , y c , T b , T c ) =

that is, as the arithmetic average of a technology c-based Luenberger indicator and a technology b- based Luenberger indicator. Input- oriented and output- oriented Luenberger indicators are generated by setting g y = 0 or g x = 0. The nomenclature indicator, as opposed to index, has been adopted to draw a distinction between difference-based measures like the Luenberger indicator and ratio- based measures like the Malmquist and Hicks-Moorsteen indices. For theoretical comparisons of ratio- based and difference- based measures, see Chambers (1998, 2002) and Diewert (2005). See Boussemart, Briec, Kerstens, and Poutineau (2003) for empirical comparisons of the two approaches to productivity measurement. Analogously to the multiplicative decomposition of the Malmquist index (4.24)– (4.27), the Luenberger indicator (4.35) can be additively decomposed as follows:

Π L (x b , x c , y b , y c , T b , T c ) = [D(x b , y b , T b , g ) − D(x c , y c , T c , g )] 1 + (D(x c , y c , T c , g ) − D(x c , y c , T b , g ) (4.36) 2 + D(x b , y b , T c , g ) − D(x b , y b , T b , g )),

where the first term (in brackets) is efficiency change and the second term is technological change. This indicator satisfies properties analogous to those of the hyperbolic index. Briec (1997) proposed a variation of the directional distance function by using the definition of D with the direction g = 〈 x , y 〉 and thereby specifying a specific distance

174 R. Robert Russell measure rather than a class of measures parameterized by g . The proportional directional distance function, defined by

DP (x , y , T ) = max λ | (1 − λ)x , (1 + λ) y ∈T ,

{

}

maps into the [0, 1] interval. Replacing the directional distance function in (4.36) with this function yields a proportional Luenberger productivity indicator. Note that this indicator, roughly speaking, is an additive analog of the hyperbolic index.

4.7. Dual Productivity Indices The Farrell technological efficiency index lies at the core of the decomposition of productivity indices in section 4.5. Farrell, however, was interested in a broader concept, encompassing allocative as well as technological efficiency. While the latter notion is a measure of the excess cost (above the minimum) attributable to operating below the production frontier (above the isoquant), the former is a measure of the extra cost accrued by operating at an inefficient point on the frontier, given input prices. Thus, examination of allocative efficiency, unlike technological efficiency, necessarily entails information about prices and economic units’ adjustments to price changes. A thorough examination of economic efficiency would expand considerably the scope of this chapter beyond the common notion of productivity as a technological phenomenon. Nevertheless, the recent introduction of notions of dual productivity indices suggests that a brief discussion is called for. The cost function, C : R n++ × R m+ × → R + , dual to the input distance function, is defined by

C( p, y , T ) = min x { p ⋅ x | x , y ∈T } = min x { p ⋅ x | DI (x , y , T ) ≥ 1},

where, of course, p is the input price vector. Under the CCD assumption of technological efficiency, the obvious candidate for a cost-based productivity index is simply the ratio of minimal costs in two situations normalized on particular output and price vectors: C( y t , pt , T b ) / C( y t , pt , T c ), t = b or c, and of course the geometric average of the two. Assume the possibility of inefficient production. Cost efficiency is defined and decomposed into allocative and technological efficiency as follows:29

CE( p, x , y , T ) :=

C( p, y , T ) C( p, y , T ) 1 = ⋅ p⋅ x p ⋅ x / DI (x , y , T ) DI (x , y , T ) (4.37) =: AEI ( p, x , y , T ) ⋅ EI (x , y , T ).

Theoretical Productivity Indices 175 Recall that the Malmquist “ideal” (input-oriented) efficiency index (equation (4.15) can be written as E (x c , y c , T b ) E (x c , y c , T c ) Π IMI (x b , x c , y b , y c , T b , T c ) = I b b b I b b c EI (x , y , T ) EI (x , y , T )

1/ 2

. (4.38)

The dual to this index, first formulated by Balk (1998, 67–70) (and recently re-examined by Zelenyuk 2006), is the Malmquist ideal cost-based productivity index, a geometric average of cost-based productivity indices normalized on technologies b and c: CE( pc , x c , y c , T b ) CE( pc , x c , y c , T c ) CΠ ( p , p , x , x , y , y , T , T ) = CE( pb , x b , y b , T b ) CE(x b , y b , pb , T c ) M

b

c

b

c

b

c

b

1/ 2

c

(4.39)

=: (CΠbM ( pb , pc , x b , x c , y b , y c , T b ) CΠ cM ( pb , pc , x b , x c , y b , y c , T c ))1/2 . To obtain some intuition about this index, consider the case of no technological or allocative inefficiency in either period. In this case, CE( pb , x b , y b , T b ) = 1 and the cost- based productivity index normalized on technology b simplifies to

CΠbM ( pb , pc , x b , x c , y b , y c , T b ) = CE( pc , x c , y c , T b ) =

C( pc , y c , T b ) . pc ⋅ x c

If T b ⊂ T c and Γ(T b ) ∩ Γ(T c ) = ∅ (reflecting global technological improvement), CE( pc , x c , y c , T b ) > 1, indicating increased cost- based productivity (under the assumption of technological efficiency). Also, if there is allocative or technological inefficiency in period b, CE( pb , x b , y b , T b ) < 1, providing additional fillip to the increase in measured productivity. The technology-c based index CΠ cM ( pc , pb , x b , x c , y b , y c , T c ) can be similarly deconstructed. Revenue-based “Malmquist” indices are analogously derived using the revenue function (dual to the output distance function),

R(r , x , T ) = max y {r ⋅ y | x , y ∈T } = max y {r ⋅ y | DO (x , y , T ) ≤ 1},

(where r ∈R m+ + is the output price vector) and revenue-efficiency indices analogous to the cost-efficiency indices (4.37) (see Balk 1998, 107–112; Zelenyuk 2006 for details).

4.8. Aggregation of Productivity Indices Productivity measurement is carried out at many levels of aggregation—for example, at the firm level, the industry/sector level, and the economy-wide level. Some efforts

176 R. Robert Russell have been made to establish conditions for consistent aggregation across these units of observation. Index the individual economic units by k = 1, …, K , so that the profiles of input vectors, output vectors, and technologies in period t are 〈 x1t , …, x Kt 〉, 〈 y1t , …, y Kt 〉 , and 〈T1t , …, TKt 〉. The aggregate technology in period t is the Minkowski sum, T t = Σ kTkt . Denote the aggregate production vector in period t by x t , y t := Σ k x kt , Σ k y kt ∈T t . The strongest form of consistent aggregation, often referred to as “exact aggregation,” is the existence of a productivity index Π and an aggregation rule Ξ such that

Π(x b , x c , y b , y c , T b , T c ) = Ξ(Π(x1b , x1c , y1b , y1c , T1b , T1c ), …, Π(x Kb , x Kc , y Kb , y Kc , TKb , TKc )).

(4.40)

To my knowledge, sufficient conditions (restrictions on the productivity index Π and the aggregation rule Ξ) for this strong form of aggregation have not been proved, and given related results in Blackorby and Russell (1999), tolerable sufficient conditions are highly unlikely to exist: they showed that, substituting efficiency indices for productivity indices, (4.40) holds only if production functions are highly linear (entailing linear isoquants and linear production possibility surfaces) and congruent across economic units. Limited aggregation results have been established for dual productivity indices, owing to the well-known (Koopmans 1957) principle of interchangeability of set summation and optimization: for example,

C p, ∑ Lk ( y k ) = ∑ C( p, Lk ( y k )). k k

This fact, combined with (4.37), yields

p, x , Lk ( y k ) = CE p, L( y k ) × S k p, x k . (4.41) CE ∑k ∑ k

(

)

(

)

where S k ( p, x k ) = p ⋅ x k /p ⋅ Σ k x k is the kth unit’s share of aggregate cost. A suitable elaboration of (4.41) to encompass multiple periods and substitution into (4.39) yields a cost- based aggregate productivity index. While this aggregation result has some power, since the aggregate index depends only on the aggregate technology, rather than the profile of technologies, it does not satisfy aggregation consistency, since aggregate cost efficiency, and hence aggregate productivity, depends on the entire distribution of both input and output quantities. Given the Koopmans interchangeability principle, this exercise yields little more than a simple summation: the aggregate economic unit really plays no specific role in the analysis.30

Theoretical Productivity Indices 177

4.9. Concluding Remarks: Malmquist versus Hicks-Moorsteen Redux The Malmquist and Hicks- Moorsteen indices, developed respectively by Caves, Christensen, and Diewert (1982) and Bjurek (1996), persist to this day as the foundation of much of the pure theory of technology-based productivity indices. While the building blocks for both are Malmquist-Shephard distance functions, the two edifices are conceptually very different and give the same answer only under very restrictive conditions. Consequently, there has been some discussion (and even a little controversy) in the literature about which is the appropriate measure—or even which is the one correct measure—of productivity change. Some of the discussion is almost semantic, revolving around the “appropriate” definition of productivity. On the one hand, some make the valid point that the common understanding of productivity is total-factor productivity— an aggregate output quantity divided by aggregate input quantity—in which case a productivity index should be an output index divided by an input index.31 The Hicks- Moorsteen index has this property and, unlike the Malmquist index, incorporates the effects of returns to scale (as well as input and output mix effects) into the measurement of productivity. Consider, on the other hand, a comparison of the “productivity” of a firm in two periods b and c, and suppose that the production possibility surface of a firm at time c is above that at time b at all input quantities but, owing to diminishing returns to scale and a lower output in period c, total factor productivity at c is lower than that at b. It seems natural to say that the firm is more productive at time c, and that the lower total factor productivity at time c, when prices are different, is simply the result of optimizing behavior and hence does not reflect lower productivity of the firm. (If the technology at time b is not a proper subset of that at c, it is natural to compare the two in the neighborhood of production in period b or c, or at the mean of the two.) It seems to me that there is not a “correct” choice between the Malmquist and Hicks- Moorsteen indices: the choice may depend on the context of the productivity question (and the proclivities of the researcher), and the important thing is to be cognizant of the differences and take them into account in analyses, particularly in the interpretation of results. From a purely axiomatic point of view, however, the Hicks-Moorsteen index would seem to have the upper hand, since it satisfies the proportionality condition (P) in addition to the conditions satisfied by the Malmquist index. (Both fail the transitivity test (T).32) While it is tempting to differentiate between the Malmquist index and the Hicks- Moorsteen index by renaming the former as an index of technological change, reserving the term productivity index for the latter, this would be misleading when we take into account

178 R. Robert Russell the possibility of technologically inefficient production (production below the technological frontier): in this case, the Malmquist index incorporates the effects of a change in technological efficiency as well as technological change (as does the Hicks-Moorsteen index). While Malmquist indices decompose cleanly into technological and efficiency components, decomposition of the Hicks-Moorsteen index and further decomposition of the Malmquist index (into scale and mix effects) have been fraught with difficulties, though creative efforts have generated some interesting results. Another nice attribute of Malmquist indices is that they can be extended quite naturally to measurement in the full 〈input, output 〉 space. (The Hicks-Moorsteen index, by way of contrast, is a hybrid of separate measurements in input space and in output space.) The most natural extension is the hyperbolic index of technological change, a fairly straightforward generalization of the hyperbolic efficiency index. For reasons that escape me, this index has not flourished. In fact, several other technological efficiency indices in 〈input, output 〉 space could be used to construct technology-based productivity indices.33 As noted earlier, the dominant nonradial productivity measure is based on the directional distance function. The DDF is actually a class of functions parameterized by the directional vector g . The discretion that this parameterization yields to the researcher is a double-edged sword. On the one hand, it opens up the possibility of incorporating a weighting scheme for inputs and outputs that reflects legitimate policy objectives. On the other hand, too much discretion can impose an unwanted burden on the researcher or, heaven forbid, can make productivity measurement subject to the specious whims of the researcher. Empirical researchers have commonly disposed of this burden by setting the direction g equal to the unit vector. Although the DDF formally satisfies unit invariance (UI), when conjoined with this convention in the selection of g , it violates (UI) in practice. Two other strands of research on theoretical productivity indices, dual indices and aggregation of indices, in my view have not yet led to important theoretical breakthroughs. Technological productivity is an important component of economic efficiency, but I think it is useful to maintain a distinction between purely internal productivity considerations and market-oriented considerations. It is possible to derive quasi-aggregation results if some (optimal) reallocation of inputs or outputs among production units is incorporated into the analysis, but this approach really circumvents the aggregation problem by effectively converting the separate decision-making units into single decisions-making units. To summarize, it seems to me that theoretical developments since the trailblazing CCD and Bjurek papers, while impressive, are not as important as the rich set of applications based on their indices. These applications are described in several chapters of this volume.

Notes 1. The chapter was substantially improved by a thorough perusal of a first draft by Bert Balk and Knox Lovell. It has also benefited from comments by Chris O’Donnell, Rolf Färe, Valentin Zelenyuk, and Oleg Badunenko.

Theoretical Productivity Indices 179 2. Allowing for more than two inputs is inconsequential; eschewing Solow’s assumption of first-degree homogeneity (constant returns to scale) of the production function is not inconsequential but is nevertheless innocuous for most of our discussion. 3. A := B means the relation defines A , and A =: B means the relation defines B . 4. Solow’s assumption of first-degree homogeneity of the production function is not needed for any of the foregoing concepts to be well defined, but without it (4.6) and (4.7) fail to satisfy the proportionality property described in subsection 4.2.5. (The Hicks-Moorsteen generalization of the Solow index does satisfy this property, but the Malmquist index does not unless the technology satisfies constant returns to scale.) 5. And (4.2) as well. 6. Nor are they appropriate for productivity measurement in input or output space in the presence of “bads” like pollution; see Chapter 8 of this volume. 7. This problem can be dealt with in other ways (see footnote 12 in Russell 1998), but given the unimportance of the shutdown condition, I think the simple approach of purging the null vector from the domain is best. 8. It is not standard in the literature to make explicit the dependence of the input-requirement and production-possibility sets on the specification of the technology, but it is convenient to do so when we contemplate productivity changes in the face of changes in the technology over time or differences of the technology among production units. 9. Mathematicians refer to it as the “gauge function”; “distance function” is a distinctive (and different) concept in real analysis. 10. The input distance function in the Solow setup is given by DIS (x , y ) = max{λ > 0 | y ≤ f (x / λ)} . If f is homogenous of degree 1, DIS (x , y ) = f (x )/y . 11. The output distance function in the Solow setup is given by DOS (x , y , T ) = min{λ > 0 | y / λ ≤ f (x )} = y /f (x ).

12. For proofs, discussions, and illustrations of the properties of input and output distance functions, see Färe and Primont (1995) and Russell (1998). 13. For thorough examinations of properties of productivity indices, see Diewert (1992b) and Balk (1998). 14. This condition is equivalent to circularity, Π(x b , x c , y b , y c ) ⋅ Π(x c , x d , y c , y d ) ⋅ Π(x d , x b , y d , y b ) = 1 , if (and only if) the identity condition holds, but the two conditions are typically used interchangeably under the implicit assumption that the latter condition holds. 15. Dependence on economic variables, like prices, is briefly discussed in subsection 4.7. 16. Comprehension of the concepts discussed throughout this chapter would be enhanced by drawing the diagrams that take up too much space to include in this survey. For diagrammatic expositions of many of these concepts, see Russell (1998). 17. I follow convention in normalizing on either the base-period or current-period technology, but the Malmquist indices could in principle be defined with respect to any technology in . Cf. note 21. 18. Use note 11 to obtain DO (x c , y c , T b ) = y c /f b (x c ) . 19. Use note 11 to obtain 1/ DO (x b , y b , T c ) = f c (x b )/y b . 20. Use note 10 to obtain 1/ DI (x c , y c , T b ) = y c /f b (x c ) . 21. See Färe and Grosskopf (1996). Input-oriented (output-oriented) Malmquist indices satisfy (P) if the distance function is homogeneous of degree –1 in output quantities (homogeneous of degree –1 in input quantities). To get around the nontransitivity problem, Berg, Førsund, and Jansen (1992) and Pastor and Lovell (2005), following a suggestion of Diewert (1987), propose indices defined on a particular technology (the first-period

180 R. Robert Russell technology or the union of all technologies). This approach is critiqued by Balk and Althin (1996), who then propose a more elaborate construction to impose transitivity. 22. But also output mix effects; see Balk (2001) and the discussion of the decomposition of a productivity change into its separate components in section 4.5. 23. Färe and Primont (1995, 69–72) showed that this property, which they called inverse homotheticity, is equivalent to DO (x , y , T ) = DO (x , y , T ) / Ψ(DI (x , y , T ) for an arbitrary 〈 x , y 〉 ∈R n++m and for some strictly monotonic function Ψ . 24. See Färe, Grosskopf, Norris, and Zhang (1994), Ray and Desli (1997), Grifell-Tatjé and Lovell (1999), and Wheelock and Wilson (1999). These contributions are reviewed by Balk (2001), Grosskopf (2003), and Zofio (2007). 25. Note that, since SO (x , y , T ) is homogeneous of degree zero in y , its value depends on y only through the location of the ray through y . 26. Bert Balk points out (in a private communication) that the numerators can also be characterized as dual input indices. 27. Note that the Hicks-Moorsteen index does not measure productivity change along a stipulated path in graph space; rather, it is an amalgamation of separate measurements along radial paths to a frontier in input space and output space. 28. See also Chambers, Chung, and Färe (1996), Chambers, Färe, and Grosskopf (1996), Färe and Grosskopf (1996), Balk (1998), and Chambers (1998, 2002). A precursor of this literature is Diewert (1983), where a directional distance function is introduced to measure waste in production. 29. Note that x / DI (x , y , T ) ∈ I ( y ) is the maximally contracted input vector. 30. Weaker aggregation conditions have been explored by Zelenyuk (2006) and Balk (2016). 31. This position has been most ardently argued by O’Donnell (2012), who goes on to develop an array of such (“multiplicatively complete”) indices and proposed decompositions. In contrast to most of the theoretical literature on multiple-output productivity indices, however, O’Donnell’s notion of total-factor productivity indices comprises “mechanistic” aggregate outputs and inputs—that is, aggregates that are not necessarily integral to the specification of the technology. 32. But see note 21. 33. See Russell and Schworm (2011) for descriptions and evaluations of these efficiency indices.

References Balk, B. M. 1998. Industrial Price, Quantity, and Productivity Indices: The Micro-Economic Theory and an Application. Boston: Kluwer Academic. Balk, B. M. 2001. “Scale Efficiency and Productivity Change.” Journal of Productivity Analysis 15: 159–183. Balk, B. M. 2016. “Various Approaches to the Aggregation of Economic Productivity Indices.” Pacific Economic Review 21: 445–463. Balk, B. M., and R. Althin. 1996. “A New, Transitive Productivity Index.” Journal of Productivity Analysis 7: 19–27. Berg, A., F. R. Førsund, and E. S. Jansen. 1992. “Malmquist Indexes of Productivity Growth during the Deregulation of Norwegian Banking.” Scandinavian Journal of Economics 94 (Supplement): S211–S228.

Theoretical Productivity Indices 181 Bjurek, H. 1996. “The Malmquist Total Factor Productivity Index.” Scandinavian Journal of Economics 98: 303–313. Blackorby, C., and R. R. Russell. 1999. “Aggregation of Efficiency Indexes.” Journal of Productivity Analysis 12: 5–20. Boussemart, J.-P., W. Briec, K. Kerstens, and J.-C. Poutineau. 2003. “Luenberger and Malmquist Productivity Indices: Theoretical Comparisions and Empirical Illustration.” Bulletin of Economic Research 55: 391–405. Briec, W. 1997. “A Graph-Type Extension of Farrell Technical Efficiency Measure.” Journal of Productivity Analysis 8: 95–110. Caves, D. W., L. R. Christensen, and W. E. Diewert. 1982. “The Economic Theory of Index Numbers and the Measurement of Input, Output, and Productivity.” Econometrica 50: 1393–1414. Chambers, R. G. 1996. “A New Look at Input, Output, Technical Change and Productivity Measurement.” University of Maryland Working Paper. Chambers, R. G. 1998. “Input and Output Indicators.” In Index Numbers: Essays in Honor of Sten Malmquist, edited by R. Färe, S. Grosskopf, and R. R. Russell, 241–271. Boston: Kluwer Academic. Chambers, R. G. 2002. “Exact Nonradial Input, Output, and Productivity Measurement.” Economic Theory 20: 751–765. Chambers, R. G., Y. Chung, and R. Färe. 1996. “Benefit and Distance Functions.” Journal of Economic Theory 70: 407–419. Chambers, R., R. Färe, and S. Grosskopf. 1996. “Productivity Growth in APEC Countries.” Pacific Economic Review 1: 181–190. Debreu, G. 1951. “The Coefficient of Resource Utilization.” Econometrica 19: 273–292. Diewert, W. E. 1983. “The Measurement of Waste within the Production Sector of an Open Economy.” Scandinavian Journal of Economics 85: 159–179. Diewert, W. E. 1987. “Index Numbers.” In The New Palgrave Dictionary of Economics, edited by J. Eatwell, M. Milgate, and P. Newman, 767–780. London: Palgrave Macmillan. Diewert, W. E. 1992a. “Fisher Ideal Output, Input, and Productivity Indexes Revisited.” Journal of Productivity Analysis 3: 211–248. Diewert, W. E. 1992b. “The Measurement of Productivity.” Bulletin of Economic Research 44: 163–198. Diewert, W. E. 2005. “Index Number Theory Using Differences Rather Than Ratios.” American Journal of Economics and Sociology 64: 311–360. Färe, R., and S. Grosskopf. 1996. Intertemporal Production Frontiers: With Dynamic DEA. Boston: Kluwer Academic. Färe, R., S. Grosskopf, and C. A. K. Lovell. 1985. The Measurement of Efficiency of Production. Boston: Kluwer-Nijhoff. Färe, R., S. Grosskopf, M. Norris, and Z. Zhang. 1994. “Productivity Growth, Technical Progress, and Efficiency Change in Industrialized Countries.” American Economic Review 84: 66–83. Färe, R., S. Grosskopf, and P. Roos. 1996. “On Two Definitions of Productivity.” Economics Letters 53: 269–274. Färe, R., S. Grosskopf, and R. R. Russell (eds.). 1998. Index Numbers: Essays in Honor of Sten Malmquist. Boston: Kluwer Academic. Färe, R., and D. Primont. 1995. Multi-Output Production and Duality: Theory and Applications. Boston: Kluwer Academic. Farrell, M. J. 1957. “The Measurement of Productive Efficiency.” Journal of the Royal Statistical Society, Series A 120: 253–290.

182 R. Robert Russell Grifell-Tatjé, E., and C. A. K. Lovell. 1995. “A Note on the Malmquist Productivity Index.” Economics Letters 47: 169–175. Grifell-Tatjé, E., and C. A. K. Lovell. 1999. “A Generalized Malmquist Productivity Index.” Sociedad de Estadística e Investigación Operativa TOP 7: 81–101. Grosskopf, S. 2003. “Some Remarks on Productivity and Its Decompositions.” Journal of Productivity Analysis 20: 459–474. Hicks, J. R. 1961. “Measurement of Capital in Relation to the Measurement of Economic Aggregates.” In The Theory of Capital, edited by F. A. Lutz and D. C. Hague, 18–31. London: Macmillan. Koopmans, T. J. 1957. Three Essays on the State of Economic Analysis. New York: McGraw-Hill. Lovell, C. A. K. 2003. “The Decomposition of Malmquist Productivity Indexes.” Journal of Productivity Analysis 20: 437–458. Luenberger, D. G. 1992. “Benefit Functions and Duality.” Journal of Mathematical Economics 21: 115–145. Malmquist, S. 1953. “Index Numbers and Indifference Curves.” Trabajos de Estatistica 4: 209–242. Moorsteen, R. H. 1961. “On Measuring Productive Potential and Relative Efficiency.” Quarterly Journal of Economics 75: 451–467. O’Donnell, C. 2012. “An Aggregate Quantity Framework for Measuring and Decomposing Productivity Change.” Journal of Productivity Analysis 38: 255–272. Pastor, J. T., and C. A. K. Lovell. 2005. “A Global Malmquist Productivity Index.” Economics Letters 88: 266–271. Peyrache, A. 2014. “Hicks–Moorsteen Versus Malmquist: A Connection by Means of a Radial Productivity Index.” Journal of Productivity Analysis 41: 187–200. Ray, S. C., and E. Desli. 1997. “Productivity Growth, Technical Progress, and Efficiency Change in Industrialized Countries: Comment.” American Economic Review 87: 1033–1039. Russell, R. R. 1998. “Distance Functions in Consumer and Producer Theory.” In Index Numbers: Essays in Honor of Sten Malmquist, edited by R. Färe, S. Grosskopf, and R. R. Russell, 7–90. Boston: Kluwer Academic. Russell, R. R., and W. Schworm. 2011. “Properties of Inefficiency Indexes on 〈Input, Output〉 Space.” Journal of Productivity Analysis 36: 143–156. Shephard, R. W. 1953. Cost and Production Functions. Princeton, NJ: Princeton University Press. Solow, R. M. 1957. “Technical Change and the Aggregate Production Function.” Review of Economics and Statistics 39: 312–320. Wheelock, D. C., and P. W. Wilson. 1999. “Technical Progress, Inefficiency, and Productivity Changes in U.S. Banking, 1984–1993.” Journal of Money, Credit, and Banking 31: 212–234. Zelenyuk, V. 2006. “Aggregation of Malmquist Productivity Indexes.” European Journal of Operational Research 174: 1076–1086. Zofio, J. L. 2007. “Malmquist Productivity Index Decompositions: A Unifying Framework.” Applied Economics 39: 2371–2387. Zofio, J. L., and C. A. K. Lovell. 2001. “Graph Efficiency and Productivity Measures: An Application to US Agriculture.” Applied Economics 33: 1433–1442.

Chapter 5

Dynam ic Effi c i e nc y and Produ c t i v i t y Rolf Färe, Shawna Grosskopf, Dimitris MargaritIs, and William L. Weber

5.1. Introduction The measurement of productivity change using linear programming methods has its roots in the Malmquist productivity index, introduced as a theoretical index by Caves, Christensen, and Diewert (1982), hereafter CCD. They showed that this index, defined directly on technology through distance functions, can be expressed as a Törnqvist productivity index by applying Diewert’s (1976) quadratic lemma.1 This link provided both index theoretic and production theoretic underpinnings to the Törnqvist index. The work was generalized by Diewert (1992), Balk (1993), and Färe and Grosskopf (1992) to include a link between the Malmquist productivity index and the Fisher productivity index. Nishimizu and Page (1982) provided an empirical application using parametric frontier estimation based on work by Aigner and Chu (1968), which used linear programming techniques to estimate parametric production frontiers. Later, beginning with Färe, Grosskopf, Lindgren, and Roos (1994), the Aigner and Chu method was extended to nonparametric frontier estimation. A 1949 conference at the Cowles Commission applied linear programming methods to analyze production decisions. Since then, a veritable industry has developed that employs what Koopmans (1951) termed activity analysis and is now associated with data envelopment analysis, hereafter DEA, to analyze performance, including efficiency and productivity. The link to estimation of performance using DEA methods is the distance function, which provides a description of technology that—unlike the classic Törnqvist and Fisher indices—does not require price data to provide a means of aggregating multiple inputs and outputs to arrive at the input and output quantity indices required to measure productivity, but does require an optimization procedure.

184 Färe, Grosskopf, Margaritis, and Weber It also provides a measure of technical efficiency, allowing for decompositions of the productivity indices. Most of the better-known total factor productivity (TFP) indices in this class, including Malmquist, Hicks-Moorsteen, and the Färe-Primont index recently promoted by O’Donnell (see, e.g., O’Donnell 2014), are what might be called comparative static measures of changes in productivity. Hans Bjurek (1996) proposed a Malmquist-type productivity index that was explicitly defined as ratios of distance function–based output and input quantity indices over time. A static technology evaluates producer performance given the current level of inputs used to produce current outputs, that is how firms allocate their inputs in the current period to produce their optimal output mix. Moreover, static descriptions of the technology do not capture the effects of how current input and output use affect future production possibilities, and hence the ability of firms to enhance productivity in the long run. Given a static technology where producers use a single input to produce a single output, productivity equals the amount of output per unit of input. A productivity change index then equals the ratio of productivity in one period to productivity in a past period. When producers use multiple inputs to produce multiple outputs, distance functions can be used to aggregate inputs and outputs. In this case a simple measure of total factor productivity is the ratio of an output distance function to an input distance function, and a total factor productivity change index equals the ratio of total factor productivity over time (O’Donnell 2014). We note that various static productivity change indices, such as the Malmquist and Hicks-Moorsteen indices, evaluate productivity change using mixed-period distance functions. The mixed-period distance functions measure the distance of current period outputs and inputs to an observed future frontier technology and the distance of future period outputs and inputs to the current frontier technology. However, these mixed- period distance functions do not constitute a dynamic problem in that current input and output decisions have no impact on future production possibilities. Based on work by Shephard and Färe (1980), who constructed index numbers in function spaces, this chapter extends the comparative static Malmquist productivity index to a dynamic setting that allows the reallocation of resources over time so that today’s resource decisions have an impact on tomorrow’s production.2 Thus, if today’s decisions restrict current outputs or appear to use inputs inefficiently but augment future production, static measures give a biased estimate of productivity change and might lead one to infer a spurious causal effect.3 We obtain the dynamic version of the Malmquist productivity index by applying distance functions to a dynamic production model, thereby mimicking the original approach taken by CCD in 1982. We also show how to estimate the dynamic distance functions using activity analysis, which follows the approach taken by Färe and Grosskopf (1996). We include an empirical illustration that compares the dynamic to the original static Malmquist productivity index. The basic approach is to begin with the underlying dynamic technology. A characteristic of dynamic technologies is that production decisions undertaken in one period affect production possibilities in subsequent periods. Examples of dynamic technologies

Dynamic Efficiency and Productivity 185 abound. The bank lending process typically generates a jointly produced undesirable byproduct: nonperforming loans. As banks accrue nonperforming loans, future lending opportunities are constrained unless offset by greater use of financial equity capital. In education, students typically take a sequence of courses, with knowledge gained in one course serving as an input to subsequent courses. Farmers can choose to produce grain and beef cattle as final outputs, or save some portion of the grain as seed for future crops and use some beef cattle to maintain or augment the herd size (see Färe and Whitaker in Färe and Grosskopf 1996 for an example). More generally, all firms that invest in capital expect to receive a flow of services from that capital over time, and so the timing of capital investment projects affects that flow. Generally, though, capital investment requires that some resources that might have been used to produce final outputs be diverted to produce the capital so as to enhance future production. For example, including investment as a variable to be endogenously determined as part of a dynamic optimization problem in the specification of the distance functions can be thought of as a network with intermediate products or multistage production that links to supply chain models. In fact, dynamic models based on DEA can be thought of as a type of network DEA model. This linkage allows for solving for the optimal allocation of investments over time, as for example in Bogetoft et al. (2009), which solves for the optimal mix of public and private investment over time for a panel of US states. Developed by Färe and Grosskopf (1996), a common example of a dynamic technology accounts for producers choosing between using current inputs to produce final outputs or an intermediate product—capital investment—that can be used to enhance future production. Static models that ignore the intermediate product give a biased measure of current and future performance. Using the stock of forests as an intermediate product, Kao (2013) shows that system efficiency is overestimated when dynamic links between periods are ignored.4 Nemoto and Goto (2003) and Sueyoshi and Sekitani (2005) measure cost efficiency in a single period by treating the past capital stock as a quasi-fixed input to the current period and the current-period capital stock as an output. Emrouznejad and Thanassoulis (2005) measure dynamic efficiency over a finite period, accounting for input used and output produced in each period, with the value of the capital stock in the final period treated as an output. Thus, two producers who begin with the same capital stock and use the same inputs and produce the same outputs in each period will have different dynamic efficiency scores if one producer has a larger final capital stock. More recent work has specified an infinite horizon dynamic technology where firms are assumed to minimize the present value of costs of production over time by choosing an optimal amount of capital investment and other inputs across time. These models have been estimated using the present-value Hamilton-Jacobi-Bellman equation for costs. Such models have been employed by Rungsuriyawiboon and Stefanou (2007) to examine the technical and allocative efficiencies of US electric utilities; Silva and Stefanou (2003, 2007) for Pennsylvania dairy farmers; Serra, Lansink, and Stefanou (2011) for Dutch dairy farmers; Skevas, Lansink, and Stefanou (2012) for Dutch arable farms; and Kapelko, Lanskink, and Stefanou (2014) for firms in the Spanish construction

186 Färe, Grosskopf, Margaritis, and Weber industry. Lansink, Stefanou, and Serra (2015) extend the dynamic efficiency measures estimated for Dutch dairy farmers to dynamic Luenberger productivity indicators, which are based on directional distance functions. Chen and van Dalen (2010) build a dynamic model to account for the effect of past advertising expenditures (an input) on future sales (output), whereas Ang and Lansink (2014) assess dynamic profit efficiency for Belgian dairy farms. The dynamic technology and performance measures can also be generalized to allow for identification of optimal starting points, input intensity, and stopping points, which we call the time substitution model. The time substitution model determines when a given amount of input is to be allocated across time so as to maximize output (Färe, Grosskopf, and Margaritis 2010). In these types of problems, production is optimized by producers choosing when to begin and end production and how intensively to use inputs. Technological progress makes it optimal for firms to delay production until later periods since the same inputs will produce greater outputs. In contrast, production should take place in earlier periods, when the production technology is experiencing stagnation or technological regress. At the optimum, the marginal rate of transformation between outputs and marginal rate of substitution between inputs should be the same across all periods. One might ask why technological regress—that is, implosion of the technological frontier—can occur if outputs and inputs are measured accurately over time. In regard to observed technological regress from 1965 to 1990 among countries with low capital/ labor ratios, Kumar and Russell (2002, 540) ask, “Were ‘blueprints’ lost?” and conclude instead that the nonparametric DEA frontier probably lies below the true but unobservable frontier at low capital/labor ratios. Pastor and Lovell (2005) measure productivity change by constructing a DEA frontier where historical input and output combinations are part of the current-period technology. As a consequence, their method attributes declines in output or increases in input usage as efficiency loss, rather than technological regress. However, there might still be instances in which technological regress occurs. When population and technological progress are endogenous, Aiya, Dalgaard, and Moav (2008) find that Malthusian population shocks can cause technological regress. Research scientists employed by for-profit firms sometimes hold tacit knowledge and face a trade-off between codifying their tacit knowledge into words, codes, and formulas versus their contractual obligations for creating new knowledge (Zucker, Darby, and Armstrong 2002). Regress might then occur when tacit knowledge exists but disappears before being passed on to other users following the death of the scientist or the firm. In many production technologies, desirable outputs and undesirable byproducts are jointly produced. Since the goal of producers facing regulation of emissions is to simultaneously maximize desirable outputs and minimize undesirable outputs, the Shephard radial output distance function, which seeks to increase all outputs, is inappropriate. In these cases, dynamic efficiency can be estimated using directional distance functions, which allow for asymmetric scaling of goods and bads, and dynamic productivity change can be estimated using Luenberger productivity indices, which are constructed from directional distance functions. Fukuyama and Weber (2015b, 2017) build a dynamic model of the bank production process where bank managers choose the amount

Dynamic Efficiency and Productivity 187 of excess reserves to carry over to a subsequent period in order to maximize loans and securities investments and simultaneously minimize nonperforming loans. In section 5.2 we present the static and dynamic production possibility sets that show how inputs are transformed into outputs and how current resource decisions impact future production possibilities. We also show how those production sets can be represented using activity analysis. Section 5.3 consists of two subsections that provide measures of dynamic performance. In subsection 5.3.1 the static and dynamic distance functions are defined and the dynamic Malmquist productivity index is presented and discussed. In subsection 5.3.2 we provide an overview of the time substitution model. In section 5.4 we offer an illustration of our dynamic productivity index for a panel of 26 Organisation for Economic Co-operation and Development (OECD) countries during the period 1990–2011 and compare it to the static Malmquist productivity index. Section 5.5 concludes the chapter.

5.2. Static and Dynamic Production Models In this section we introduce the production models relative to which we define the static and dynamic productivity indices. We denote inputs as x = x1 , . . . , x N ∈ℜN+ and outputs by y = y1 , . . . , y M ∈ℜ+M . Then the technology may be described by the output sets

(

(

)

{

)

}

P ( x ) = y : x can produce y , x ∈ℜN+ . (5.1)

We assume that the output sets satisfy the standard regularity conditions including free disposability of inputs and outputs, and convex, closed, and bounded output sets.5 We also assume that the output sets satisfy constant returns to scale so the productivity index can be interpreted as the ratio of an index of outputs to an index of inputs between two periods or between two producers, which, in the case of a single output and a single input, equals the ratio of average products. In addition, productivity growth measures are biased when variable returns to scale are assumed and can no longer be interpreted as the ratio of average products. For instance, when producers operate in the range of increasing returns to scale and input growth occurs, productivity change is understated; when firms operate in the range of decreasing returns to scale and input growth occurs, productivity change is overstated (Grifell-Tatje and Lovell 1995). Furthermore, we decompose dynamic productivity growth into the product of an index of efficiency change and an index of technical change. Both our static and dynamic production models consist of a sequence of output sets P τ x τ , τ = 1, . . . , T . In the static case these are not interconnected, except by the march of time, but in our very simple dynamic model they are interconnected by intermediate products. Thus at each period τ , the output vector y τ can be either used as final output f τ y for consumption in period t or as intermediate products i y τ so that y τ = fy τ + i y τ.

( )

188 Färe, Grosskopf, Margaritis, and Weber Final outputs

fy t−1

Intermediate products

i t−2 y

Exogenous inputs

fy t

fy t+1

i t y

i t−1 y

i t+1 y

Pt−1

Pt

P t+1

x t−1

xt

x t+1

Figure 5.1. The dynamic technology.

The intermediate products can be used as inputs in a future period to augment production. In Figure 5.1 we provide an illustration.6 Three production sets (P t −1 , P t , and P t +1) are illustrated in Figure 5.1. These form a directed network, that is, a network that is connected in a specific forward direction. It may also be thought of as a discrete Ramsey (1928) model, where output at, say, period t can be used either as final output (consumption) or as input into the following period’s production (investment or saving). In this simple dynamic structure, where only intermediate product links the adjacent periods, it is clear that if iy τ is zero for all periods so that all output y τ is final output fy τ , then the dynamic model collapses to the comparative static case (i.e., there is no intertemporal interaction, and today’s production decisions have no further consequences and are not affected by history). In the empirical illustration of a simple dynamic Malmquist productivity index, we will use an activity analysis or DEA formulation of the output sets. Suppose we are given a set of k = 1, …, K observations of inputs and outputs ( x k , y k ), then the static activity analysis model of technology in a single period is constructed from these data as7

P (x ) =

{ ( y , . . . , y ) : y ≤_ ∑ z y K

1

M

m

k =1 K

km

xn ≥ _ ∑z k x kn , k =1

k

,

m = 1, . . . , M , (5.2) n = 1, . . . , N ,

}

zk ≥ _ 0, k = 1, . . . , K .

Thus the technology is created from the data ( x k , y k ) , k = 1, . . . , K by forming the smallest convex cone that includes the data, hence the moniker data envelopment analysis. This cone technology is the constant returns to scale technology and is formed from the intensity variables, z k , k = 1, . . . , K such that no more output can be produced using no less input than a linear combination of the observed outputs and inputs. If the observed inputs and outputs satisfy the following conditions set out by Kemeny, Morgenstern, and Thompson (1956), (i) Σ kK=1 x kn > 0, n = 1, . . . , N , each input is used by at least one k (ii) Σ nN=1 x kn > 0, k = 1, . . . , K , each k uses at least one input, n

Dynamic Efficiency and Productivity 189 (iii) Σ kK=1 y km > 0, m = 1, . . . , M , each output is produced by at least one k (iv) Σ mM=1 y km > 0, k = 1, . . . , K , each k produces at least one output, m , then the technology satisfies the standard regularity conditions of free disposability of inputs and outputs, and convex, closed, and bounded output sets (Färe and Primont 1995). The dynamic activity analysis model is formed from the data τ x k , y kτ , k = 1, . . . , K τ , τ = 1, . . . , T . As an illustration, following Färe and Grosskopf (1996), we can write out the three-period model from Figure 5.1 as

(

)

(

) {(

P x t −1 , x t , x t +1 , i y t − 2 = f t −1 m

y

K

t −1

f

y t −1 , f y t ,

(

f

))

y t +1 + i y t +1 : (5.3)

t −1 t −1 + i ymt−1 ≤ _ ∑z kt−1 ( fy km + i y km ), m = 1, . . . , M , k =1

xnt−1 ≥ _

i

y

t −2 m

K t −1

∑z k =1

t −1 t −1 k kn

x , n = 1, . . . , N ,

K t −1

t −2 ≥ _ ∑z kt−1 (i y km ), m = 1, . . . , M , k =1

z kt−1 ≥ _ 0, k = 1, . . . , K t −1

Kt

t t y + i ymt ≤ _ ∑z kt ( fy km + i y km ), m = 1, . . . , M ,

f t m

k =1 Kt

t xnt ≥ _ ∑z kt x kn , n = 1, . . . , N ,

k =1

i

y

Kt

t −1 ≥ _ ∑z kt (i y km ), m = 1, . . . , M , k =1

t k

z ≥ _ 0, k = 1, . . . , K t

t −1 m

f t +1 m

y

i t +1 m

+y

K t +1

t +1 i t +1 ≤ _ ∑z kt+1 ( f y km + y km ), m = 1, . . . , M , k =1

K t +1

t +1 xnt+1 ≥ _ ∑z kt+1 x kn , n = 1, . . . , N , k =1

i

K t +1

t ymt ≥ _ ∑z kt+1 (i y km ), m = 1, . . . , M , k =1

z

t +1 k

}

≥ _ 0, k = 1, . . . , K t +1 .

The three-period dynamic model of P (.) gives the endogenously determined final outputs ( fy t −1 , fy t , fy t +1 ) that can be produced in periods t −1, t, and t + 1 with the exogenous inputs (x t −1 , x t , x t +1 , i y t − 2 ) and the endogenous intermediate products ( iy t −1 , iy t , iy t +1 ).8 The right-hand side of the output and input constraints form the

190 Färe, Grosskopf, Margaritis, and Weber technology by taking linear combinations of the outputs and inputs of the k = 1, . . . , K producers. The left-hand side variables are the outputs and inputs, including the intermediate products from the previous period that are feasible given the right-hand side technology. This model is consistent with that depicted in Figure 5.1, where i τ y , τ = t − 1, t , t + 1 are the intermediate products that make the model dynamic. Allowing the intensity variables, z kτ, to vary across periods τ = t − 1, t , t + 1 allows each period’s technology to vary. One can show (see Färe and Grosskopf, 1996) that if each technology P τ (.) , τ = t − 1, t , t + 1 satisfies the standard regularity conditions, then so does the associated dynamic model—for example, the three-period technology specified in the preceding. Moreover, if there are no intermediate products, i y τ, then the dynamic model simplifies to three sequential static models. In the dynamic model, increases in the intermediate products that are produced in period τ expand the output set in the subsequent period. In section 5.3 we define Shephard (1970) distance functions for the static output set represented by (5.2) and dynamic distance functions for the dynamic output sets represented by (5.3).

5.3. Static and Dynamic Performance Measures 5.3.1. Static and Dynamic Productivity Indices The classic definition of productivity (in level terms) is simply the ratio of output to input. This becomes complicated when there are multiple inputs and outputs; the challenge is how best to aggregate these. CCD (1982) appealed to a natural aggregator function, namely the distance function, to represent technology and provide a means of aggregating inputs and outputs. They considered their resulting Malmquist productivity index to be purely theoretical, since they considered the component distance functions to be unobservable. They used the Malmquist index to provide theoretical underpinnings for the empirical Törnqvist productivity index. The Malmquist productivity index proposed by CCD was defined using Shephard’s distance functions, which can be estimated using DEA and, in contrast to the Törnqvist index, do not require price data to aggregate inputs and outputs. Here we follow Färe, Grosskopf, Lindgren, and Roos (1994) and use the index they developed based on the output distance function.9 We begin with technology. Let P ( x ) be an output set. Then the associated output distance function is defined as

{

}

Do ( x , y ) = inf θ : ( y / θ) ∈P ( x ) , (5.4)

Dynamic Efficiency and Productivity 191 hence it is the largest feasible radial extension of the observed output vector y.

(

)

(

)

For the input-output vectors x t , y t and x t +1 , y t +1 , CCD define two Malmquist output-oriented productivity indices, namely

(

)

(

)

(

)

Mot = Dot x t +1 , y t +1 / Dot x t , y t (5.5)

and

(

)

Mot +1 = Dot +1 x t +1 , y t +1 / Dot +1 x t , y t

(5.6)

where the first is defined relative to the technology in period t , P t ( x ), and the second is defined relative to the t + 1 technology P t +1 ( x ). One can prove (see Färe, Grosskopf, and Roos 1998), that these two indices are equal if and only if the output sets are of the form P t ( x ) = A (t ) P ( x ) , (5.7)

which implies that technology and technical change are Hicks output-neutral. To avoid imposing this condition or having to arbitrarily choose one or the other of the preceding indices, Färe, Grosskopf, Lindgren, and Roos (1994) use the geometric mean of these indices as their definition of productivity change.10 Here we adopt that approach and define the (comparative static) Malmquist productivity index as

(

)

Mo x t , y t , x t +1 , y t +1 = ( Mot ⋅ Mot +1 )1/2 (5.8)

(

)

(

Dot x t +1 , y t +1 Dot +1 x t +1 , y t +1 = Dot x t , y t Dot +1 x t , y t

(

)

(

)

)

1/ 2

.

Färe, Grosskopf, Lindgren, and Roos (1989, 1994) also show that the index can be decomposed into an efficiency change and a technical change component, namely,

(

)

(

)

EFFCH = Dot +1 x t +1 , y t +1 / Dot x t , y t (5.9)

and

( (

) )

( (

) )

Dot x t +1 , y t +1 Dot x t , y t TECH = t +1 t +1 t +1 t +1 t t Do x , y Do x , y

1/ 2

(5.10)

so that

(

)

Mo x t , y t , x t +1 , y t +1 = EFFCH ⋅ TECH . (5.11)

192 Färe, Grosskopf, Margaritis, and Weber The distance functions for the comparative static CCD-type Malmquist index can be estimated using an activity analysis or DEA approach. For example, we may estimate the distance functions for the period t technology defined in (5.2) as

( D ( x , y )) t o

t

t

−1

= max λ (5.12)

K

t t s.t. λ y km ≤ _ ∑z kt y km , m = 1, . . . , M ,

k =1 K

t t x kn ≥ _ ∑z kt x kn , n = 1, …, N ,

k =1

zk ≥ _ 0, k = 1, . . . , K ,

and

(D (x t o

t +1

, y t +1

K

))

−1

= max λ (5.13)

t +1 t s.t. λ y km ≤ _ ∑ z kt y km , m = 1, . . . , M k =1

K

t +1 t x kn ≥ _ ∑z kt x kn , n = 1, …, N , k =1

zk ≥ _ 0, k = 1, . . . , K .

In both sets of problems, the technology is determined by the data from period t through the convex combinations formed with the activity variables z k . The first problem assesses data from that same period t, whereas the second problem assesses data from period t + 1 relative to the period t technology. By interchanging t + 1 and t in (5.13) we can have data from period t assessed relative to the period t + 1 technology, yielding the distance function Dot +1 (x t , y t ). Before turning to the dynamic Malmquist productivity index, we note that the Färe- Primont and Hicks-Moorsteen productivity indices also are constructed from adjacent- period distance functions. However, they use the distance functions to construct Malmquist quantity indices of output and input, with the ratio forming the productivity index, which follows the classic definition of productivity as the ratio of output to input. Following O’Donnell (2014), they differ slightly with respect to specification of the distance functions. The general specification for the Färe-Primont version for the quantity indices is

QI =

XI =

Do ( µ x , qit , µ z )

Do ( µ x , qks , µ z )

( D (x

(5.14)

) (5.15) ,µ ,µ )

Di xit , µ q , µ z i

ks

q

z

where QI is the output quantity index and XI the input quantity index. The z variables are included as environmental variables (although they were not included in the original Färe- Primont specification). Following O’Donnell (2014, 190) the “µ x , µ q , µ z

Dynamic Efficiency and Productivity 193 are arbitrary vectors that in most empirical applications would be representative of all inputs, outputs and time periods in the dataset (e.g., sample means).” The Hicks- Moorsteen productivity index is typically specified using fixed-period data for the inputs and z variables in the output quantity index and for the outputs and z variables in the input quantity index. The dynamic Malmquist productivity index is defined—like the static Malmquist productivity index—in terms of distance functions; in this case, however, they are dynamic distance functions. Let ∆ to x t , fy t , i y and ∆ to+1 x t +1 , fy t +1 , i y be dynamic distance functions as defined in the following, where i y denotes the initial and terminal conditions for the intermediate products. Then the dynamic Malmquist productivity index is

(

)

(

(

)

)

(

∆ to x t +1 , fy t +1 , i y ∆ to+1 x t +1 , f y t +1 , i y Ωo x , y , x , y , y = t t f t i × ∆ to+1 x t , fy t , i y ∆ o x , y , y

(

t

t +1

f t

f t +1

i

)

(

)

(

)

)

1/ 2

. (5.16)

Recall that total outputs equal the sum of final outputs and intermediate products: y = fy + i y . Whereas the static distance functions are defined in terms of total outputs, the dynamic distance functions are defined in terms of final outputs and exclude the intermediate products that are produced and included in total output. As we noted earlier, if there are no intermediate products, then this simple dynamic index simplifies to the usual Malmquist productivity index. For the dynamic technology represented by (5.3), the dynamic distance functions t ∆ o x t , fy t , i y and ∆ to+1 x t +1 , fy t +1 , i y are defined as

(

)

(

)

(

) y ) = 1/λ

∆ to x t , f y t , i y = 1 / λ *t (5.17)

∆

where λ and λ *t

*t +1

t +1 o

(x

t +1

f t +1

, y ,

i

* t +1

are the maximizers in the following problem

max ψ 1λ1 + . . . + ψ t λ t + ψ t +1λ t +1 + . . . + ψ T λT (5.18)

s.t.

(

)

λ1 f y 1 + i y 1 ∈ P 1 x 1 , i y 0

λ y + y ∈P t x t , iy t −1 t f t

λ

t +1 f t +1

y

i t

(

i t +1

t +1

+ y

∈P

(x

)

t +1

i t

)

, y

λT f y T + i y T ∈P T x T , i y T −1 .

(

τ

τ

)

In (5.18) each λ is weighted by ψ to account for the producer’s rate of time preference. If production is valued equally across all periods then one might choose, ψ τ = 1, τ = 1, . . . ,T . Alternatively, given an interest rate r, the weights might be chosen as ψ τ = 1 / (1 + r )τ −1 so that any expansion of production in later periods is valued less than an equivalent expansion in production in earlier periods. The optimization problem (5.18) chooses an optimal amount of intermediate products in each period so as to maximize

194 Färe, Grosskopf, Margaritis, and Weber the proportional expansion (λ t ) of observed final output to the frontier of final output summed over all periods, t = 1, . . . , T . The amount of intermediate product produced in period t determines the size of the output set and potential final output in period t + 1. Therefore, it might sometimes be optimal for the producer to forgo final outputs and instead produce the intermediate product so as to expand final output by a greater amount in the subsequent period. For example, in a two-period dynamic problem, suppose that the actual amounts of final output and intermediate product in t = 1 are f y1 = 100 and i 1 y = 10, respectively. Given the amounts of exogenous inputs x1 and x 2 and intermediate product from period t = 0, suppose that λ1 = 1.05 and λ 2 = 1.08, with their sum equal to 2.13. In this case, total frontier output in t = 1 would equal y1 = 1.05 × 100 + 10 = 115. If the amount of intermediate product was chosen optimally, say iy *1 = 13, then the amount that final output can be expanded in period 1 falls to λ1 = 1.02 with y1 = 1.02 × 100 + 13 = 115. However, if the increase in intermediate product produced in t = 1 expands the output set 1 2 in t = 2 such that λ 2 ≥ _ 1.11 (so that λ + λ ≥ 2.13) then the reallocation toward producing greater amounts of the intermediate product in t = 1 improves dynamic efficiency. In (5.18) there are a total of T periods, and the initial conditions include i y o as the initial intermediate product vector. Transversality conditions include i y T as the terminal intermediate product vector. We place a “ _” over the exogenous initial and terminal values of the intermediate products to distinguish them from the endogenous intermediate products that are chosen in periods t = 1, . . . , T − 1 as part of the optimization problem.11 Since our model is not stochastic, the terms exogenous and endogenous have the standard meanings of determined outside and determined within the model. To simplify the notation, let i y = (i y 0 ,i y T ). Each of the individual period technologies use intermediate products (which are variables to be solved in the optimization problem) and exogenous inputs x. While the dynamic distance functions scale only final outputs f y to the dynamic frontier, the static distance functions scale total output to the static frontier. Moreover, the dynamic distance functions ∆ to x t , f y t , i y , t = 1, . . . , T are obtained by solving the dynamic optimization problem (5.18) once, whereas the distance functions defined in the static case by (5.4) must be solved T times. As with the traditional static Malmquist productivity index, the dynamic productivity index may be decomposed into an efficiency change and a technical change component,

(

( (

∆ to+1 x t +1 , f y t +1 , i y Ω o x t , f y t , x t +1 , f y t +1 , i y = ∆ to x t , f y t , i y

(

)

)

( (

)

) (5.19)

) )

( (

) )

∆ to x t +1 , f y t +1 , i y ∆ to x t , f y t , i y × t +1 t +1 f t +1 i ∆ o x , y , y ∆ to+1 x t , f y t , i y

1/ 2

.

Dynamic Efficiency and Productivity 195 where DEFFCH =

∆ to+1 (x t +1 , f y t +1 , i y ) represents dynamic efficiency change and ∆ to (x t , fy t , i y )

∆ t (x t +1 , fy t +1 , i y ) ∆ to (x t , fy t , i y ) DTECH = t +o 1 t +1 f t +1 i t +1 t f t i ∆ (x , y , y ) ∆ (x , y , y ) o

1/ 2

represents dynamic technical change.

o

To calculate each of the components of the dynamic productivity index, two mixed- period distance functions ∆ to+1 (x t , f y t , i y ) and ∆ to (x t +1 , fy t +1 , i y ) must be derived. The mixed-period distance function ∆ to+1 (x t , f y t , i y ) measures the distance from observed quantities of final outputs in period t to the period t + 1 technology. This function is estimated as

max ψ 1λ1 + . . . + ψ t λ t + ψ t +1λ t +1 + . . . + ψ T −1λT −1 (5.20)

s.t.

(

λ

)

λ1 f y 1 + i y 1 ∈ P 2 x 2 , i y 1

λ

t +1 f

T −1 f

y

y

t +1

T −1

i

+ y i

+ y

t +1

T −1

∈ P t + 2 x t + 2 , i y t +1

(

)

∈P T x T , i y T −1 .

(

) y ) measures the distance from

(

The mixed-period distance function ∆ to x t +1 , f y t +1 , i observed quantities of final outputs in period t + 1 to the period t technology. These functions are estimated from the problem

max ψ 2 λ 2 + . . . + ψ t λ t + ψ t +1λ t +1 + . . . + ψ T λT (5.21)

s.t.

(

)

λ 2 fy 2 + i y 2 ∈P 1 x1 , i y o λ

t +1

f t +1

y

i t +1

+ y

∈P t x t , i y t −1

(

)

λ y + y ∈P T −1 x T −1 , i y T − 2 . T f T

i

T

(

)

Problems (5.20) and (5.21) are each solved once for each decision-making unit and the T −1 estimates of λ *t are obtained. From (5.20) the function ∆ to+1 x t , f y t , i y = 1 / λ *t and from (5.21) the function ∆ to x t +1 , f y t +1 , i y = 1 / λ *t +1. Next we introduce the static distance functions Dot x t , y t and Dot +1 x t +1 , y t +1 into the dynamic efficiency change component (DEFFCH ) so that we can decompose it into two dynamic parts and one static part. Typically, the static distance functions do not include intermediate products as variables to be solved for in the optimization. Instead, i intermediate products ( y) are taken to equal zero, or they are treated as given or exogenous in the optimization. To facilitate comparisons between dynamic and static cases, we change our notation slightly and describe the static distance functions in terms of

(

)

(

(

)

)

(

)

196 Färe, Grosskopf, Margaritis, and Weber final output, f y t = y t − iy t , as in the case of the dynamic distance functions. Intermediate products will be treated the same as the exogenous inputs from the dynamic distance function. That is, for the static distance functions the given inputs are x t and iy t −1, where x t includes the exogenous inputs in the dynamic technology. We substitute the fy terms for the y terms in the static distance functions and define efficiency change as

( (

∆ to+1 x t +1 , fy t +1 , i y DEFFCH = ∆ to x t , f y t , i y

)

( (

) (5.22)

) )

∆ t +1 x t +1 , fy t +1 , i y o t +1 t +1 f t +1 i t t +1 t +1 f t +1 i t Do x , y , y Do x , y , y × = t t f t i t −1 t t f t i ∆ o x , y , y Do x , y , y t t f t i t −1 Do x , y , y

(

(

)

)

( (

)

) .

The first term on the right-hand side is an index of the gain in efficiency from dynamic reallocation of the intermediate products from period t to t + 1. The numerator equals the ratio of the distance from the dynamic frontier to the static frontier in period t + 1, and the denominator equals the ratio of the distance from the dynamic frontier to the static frontier in period t. This index takes a value greater than one when there has been an improvement in dynamic reallocation, resulting in the static frontier moving closer to the dynamic frontier from t to t + 1, and takes a value less than one when the static frontier moves further away from the dynamic frontier because of greater misallocation of the intermediate products between periods. The second term equals the static change in efficiency that results from a producer catching up or falling behind the static frontier. This static change in efficiency takes a value greater than one when the producer is closer to the static frontier in t + 1 relative to the static frontier in t. The product of the two terms in (5.22) give the efficiency gain between the two periods due to the advantages of the reallocation allowed in the dynamic case relative to the static case. One disadvantage of the dynamic specification in the preceding is that the results will depend on the initial conditions and the transversality conditions; that is, if we wish to add one more year to our analysis, we would be re-estimating productivity change for the entire time period, which could change the results for the earlier periods.12 This problem is avoided in the traditional and sequential DEA comparative static productivity measures, which estimate productivity change for each pair of adjacent periods, but is present in the global Malmquist index (Pastor and Lovell 2005). One could avoid the end-point problem in a dynamic context by estimating (5.18) using only a subset of the entire sample period, say only three periods at a time in a horizon where there are T > 3 periods. The estimation would be then be done multiple times: first, data from period t = 0, 1, and 2 would be used allowing for reallocation of the intermediate product from t = 1 to t = 2; second, data from periods t = 1, 2, and 3 would be used allowing for

Dynamic Efficiency and Productivity 197 reallocation of the intermediate product from t = 2 to t = 3; and so forth, with the final estimation using data from periods t = T − 2, T −1, and T allowing for reallocation of the intermediate product from t = T −1 to T. Of course, one can use subsets of the time series longer than three periods. However, as long as the number of periods used is less than the entire sample period t = 0, . . . , T the shorter time period limits possible reallocation across time. As a practical matter, choosing a short time horizon for the dynamic specification has several advantages. First, even a short two-or three-period horizon provides information that is not available in the static model specification. Second, a two-or three-period horizon is likely being more realistic in that it limits the time frame that the decision-maker is required to optimize over. Third, a shorter horizon means that productivity estimates do not have to be continually updated as new information becomes available. An advantage to using all available periods to estimate the dynamic distance functions is that not only can one observe the pattern of dynamic productivity change and its components, but also optimal investment spending can be compared with the history of actual investment spending. This information might enhance policymakers’ ability to engage in countercyclical policy. Furthermore, the dynamic model specification can be easily extended to cases where the technology generates both desirable outputs and jointly produced undesirable outputs, such as real gross domestic product (GDP) and carbon equivalent emissions. Here, information about productivity change and the history of optimal and actual investment spending could inform policymakers about how quickly new investment spending might be expected to reduce carbon emissions. When examining resources allocated for education, knowledge of optimal and actual investments to enhance human capital might inform policymakers about the amounts of money to allocate to K–12 education, higher education, and extended learning opportunities.

5.3.2. Time Substitution In this section we briefly outline the idea of time substitution—another type of dynamic production model—as developed by Färe, Grosskopf, and Margaritis (2010). In the dynamic optimization problem (5.18) the amount of intermediate product is chosen so as to maximize the amount of final output to be produced over all periods, given exogenous inputs available in each period. In contrast, time substitution would constrain the amount of exogenous inputs to some fixed amount, say the sum of inputs over all periods, and would choose when to begin and end production and how intensively to use those inputs so as to maximize final outputs. Suppose that production can take place anywhere from period t = 1 to t = T . If producers have a limited amount of resources to use during the period, when should those resources be used in order to maximize total production? This question relates to the issue of time substitution, where producers choose not only the amounts of intermediate products to produce, but also when to begin production, τ , and when to

198 Färe, Grosskopf, Margaritis, and Weber end production, τ + Γ, so as to maximize final outputs. Färe, Grosskopf, and Margaritis (2010) showed that when there are increasing returns to scale, inputs should be used intensively in a single period, and when there are decreasing returns to scale, inputs should be spread over as many periods as possible, with the proportion of inputs to use in each period equal to 1/T. Furthermore, when technological progress occurs, it is optimal to delay production, and when technological regress occurs, it is optimal to produce sooner. Let xn = ΣTt =1 xnt, n = 1, . . . , N represent the amount of input available to use over all periods. The time substitution problem can be written as

max ψ τ λ τ + . . . + ψ τ + Γ λ τ + Γ (5.23)

s.t. λ τ f y τ + i y τ ∈P τ x τ , i y τ −1

(

λ

λ

τ +1 f

τ+Γ f

y

y

τ+Γ

τ +1

τ+Γ

∑x t =τ

t n

i

+ y i

+ y

τ +1

τ+Γ

∈P

τ +1

(x

)

τ +1 i

)

, yτ

∈P τ + Γ x τ + Γ , i y τ + Γ −1

(

)

≤ xn , n = 1, . . . , N .

The choice variables in the time substitution problem (5.23) are when to begin production, τ , when to end production, τ + Γ, the amounts of inputs to use in each period, xt , t = τ, . . . , τ + Γ , and the amount of intermediate product (iy t ) to produce in periods t = τ + 1, . . . , τ + Γ, with the objective of maximizing the proportional expansion of final outputs summed over all production periods. We note that τ = 0 and Γ = T are possible solutions to (5.23), but other solutions having shorter horizons with τ > 0 and Γ < T are also possible. As in the dynamic optimization problem (5.18), the producer faces a trade-off between expanding final outputs or intermediate products. Producing more intermediate products in the current period expands the output set in the subsequent period, but at a cost; fewer final outputs can be produced in the current period. The time substitution problem is related to the Shephard indirect output distance function. Given input prices, the indirect output distance function allows inputs to be reallocated as long as they satisfy a cost constraint consistent with the status quo. Since the status quo inputs are feasible, but not necessarily optimal, a greater proportional expansion in outputs is possible relative to Do (x , y )−1. However, unlike the Shephard indirect output distance function, the time substitution problem (5.23) does not require one to know input prices or have a targeted cost constraint, only that the reallocated inputs be no greater than the sum of inputs across time. In terms of productivity, the sum over time of the potential final outputs produced with time substitution will be greater than the same sum without time substitution, but the total amount of inputs will be the same. However, with time substitution some periods will see more inputs used than the status

Dynamic Efficiency and Productivity 199 quo and some periods will see lower levels of potential output produced, so that a productivity change index might be either greater or less than a static productivity index in some periods. Färe, Grosskopf, Margaritis, and Weber (2012) use time substitution to determine when countries should reduce emissions of carbon dioxide to be in compliance with the Kyoto Treaty. Fukuyama, Weber, and Xia (2016) examine the allocation of National Science Foundation funds for nanobiotechnology research between universities and across time in another application of time substitution. Fukuyama and Weber (2015a) examine bank production decisions using time substitution. They estimate a quadratic directional output distance function and then use the parameter estimates to calculate when it would be optimal for banks to reduce nonperforming loans so long as performing loans and securities investments were not reduced and costs of production were no greater than actual costs. Färe, Grosskopf, Margaritis, and Weber (2015) use the time substitution model to examine the European Union Stability and Growth Pact and how countries in the Union might optimally meet their government budget obligations, highlighting the fact that strict adherence to rigid fiscal rules might entail large costs in terms of lost output, as evidenced by the case of Greece.

5.4. Empirical Illustration of Static and Dynamic Performance To illustrate our method, we use pooled data taken from the Penn World Tables for 33 OECD countries during the period 1990–2011.13 We focus on estimating dynamic productivity change, rather than the time substitution problem. Countries produce total output equal to real GDP. We ignore discounting and assume the weights for each period are ψ t = 1. Total output (y) consists of an intermediate product (iy)—real investment spending—and a final output ( fy) equal to real GDP less investment spending. Labor is a given input (x) in each period and is not reallocated between periods. In period t = 1 , which corresponds to 1991, the producer has access to a given amount of investment spending from period 0, and this given investment spending (i y 0 = i y 1990) is an input to the period t = 1 (1991) technology. Thus, in period t = 1 there are two given inputs—labor and prior investment spending—that are used to produce a single total output, which equals the sum of final output and current period investment spending. In t = 1 (1991) to t = T − 1 (2010) the amount of investment is chosen as part of the optimization. This chosen investment spending becomes an input in periods t = 2 (1992) to t = T (2011) and is used, along with labor, to produce total output, which equals the sum of final output and current investment spending. In period T (2011) investment spending is fixed to

200 Färe, Grosskopf, Margaritis, and Weber satisfy the transversality condition: i y T = i y 2011. Using DEA, the optimization problem is to calculate the dynamic distance functions for country k ′ in the form:

max λ1 + . . . + λT (5.24)

s.t. λ1 f y1k ′ + i y1 ≤ ∑ z 1k y1k

K

k =1 K

y k0′ ≥ ∑ z 1k i y k0

i

k =1 K

x1k ′ ≥ ∑ z 1k x1k

z 1k ≥ 0, k = 1, . . . , K , t = 1,

k =1

K

λ t f y kt ′ + i y t ≤ ∑ z kt y kt

k =1

i t −1

y

K

≥ ∑ z kt i y kt−1 k =1 K

x k2′ ≥ ∑ z kt x kt

z kt ≥ 0, k = 1, . . . , K , t = 2, . . . , T − 1,

k =1

K

λT fy Tk ′ + i y kT′ ≤ ∑ z Tk y Tk

k =1

i T −1

y

K

≥ ∑ z Tk i y Tk −1 k =1 K

x Tk ′ ≥ ∑ z Tk x kT

z Tk ≥ 0, k = 1, . . . , K , t = T .

k =1

The LP problem (5.24) is solved once for each of the 33 countries in our sample. In (5.24) the choice variables are λ t , t = 1, . . . , T , which represent the maximum feasible expansion of final outputs in each period. The solution to (5.24) provides estimates of λ t for each period t = 1, . . . , T corresponding to 1991 to 2011. Dynamic efficiency equals ∆ to x t , f y t , i y 0 = 1 / λ *t , which is a vector of length T. In addition, the intensity variables, z kt , t = 1, . . . , T and the amount of optimal investment (i y t ) to undertake in periods t = 1 to t = T −1 are also chosen as part of the optimization. Investment in period t = 0, i y k0’ enters as an input to the period t = 1 technology but is given. Similarly, investment in period t = T , i y kT’ enters the output equation but is taken as given. Problem (5.24) uses data

(

)

Dynamic Efficiency and Productivity 201 from periods t = 0, 1, . . . , T corresponding to 1990 to 2011. We solve (5.24) one time for each of the K = 33 countries in our sample. To estimate dynamic productivity change, we solve two mixed-p eriod problems that are the DEA representations of (5.20) and (5.21). The first mixed-p eriod problem estimates the distance from the observed outputs in period t to the technological frontier defined by the outputs and inputs in periods t + 1. This problem is written as

max λ1 + . . . + λT −1 (5.25)

s.t. λ1 fy1k ′ + i y1 ≤ ∑ z k2 y k2

K

k =1

i

K

y k0′ ≥ ∑ z k2 i y1k k =1 K

x1k ′ ≥ ∑ z k2 x k2

z k2 ≥ 0, k = 1, . . . , K , t = 1,

k =1

K

λ t fy kt ′ + i y t ≤ ∑ z kt+1 y kt+1

k =1

i t −1

y

K

≥ ∑ z kt+1 i y kt k =1 K

x kt ′ ≥ ∑ z kt+1 x kt+1

t +1 k

k =1

z

≥ 0, k = 1, . . . , K ,

K

λT −1 f y Tk ′−1 + i y kT′−1 ≤ ∑ z Tk y Tk

k =1

i T −2

y

K

≥ ∑ z Tk i y Tk −1 k =1 K

x Tk ′−1 ≥ ∑ z Tk x kT

k =1

T k

z ≥ 0, k = 1, . . . , K .

The mixed- period efficiency found as part of the solution from (5.25) equals t f t i ∆ x , y , y = 1 / λ *t . The second mixed-period problem estimates the distance from the observed outputs in period t + 1 to the technological frontier defined by the outputs and inputs in periods t. This problem is written as t +1 o

(

)

202 Färe, Grosskopf, Margaritis, and Weber

max λ 2 + . . . + λT (5.26)

s.t. λ 2 fy k2′ + i y 2 ≤ ∑ z 1k y1k

K

k =1

i

K

y k1′ ≥ ∑ z 1k i y k0 k =1 K

x k2′ ≥ ∑ z 1k x1k

z 1k ≥ 0, k = 1, . . . , K ,

k =1

K

λ t +1 fy kt+′ 1 + iy kt+1 ≤ ∑ z kt y kt

k =1 K

y ≥ ∑ z kt i y kt−1

i t

k =1 K

x kt+′ 1 ≥ ∑ z kt x kt

k =1

t k

z ≥ 0, k = 1, . . . , K ,

λT f y Tk ′ + i y kT ≤ ∑ z Tk −1 y Tk −1

K

k =1

i

K

y T −1 ≥ ∑ z Tk −1 i y Tk − 2 k =1 K

x Tk ′ ≥ ∑ z Tk −1 x Tk −1

T −1 k

k =1

z

≥ 0, k = 1, . . . , K .

The mixed- period efficiency found as part of the solution from (5.26) equals ∆ to x t +1 , f y t +1 , i y = 1 / λ *t +1. Table 5.1 reports the geometric means for the dynamic and static efficiency estimates by year and the number of countries defining the frontier. In our static model specification, each country produces a final output ( fy t ) equal to real GDP less real investment spending using two inputs: labor (x t ) and prior investment spending (i y t −1). Average efficiency over all years for the dynamic model ∆ t x , fy , i y t −1 = 0.549 is less than average efficiency for the static model Dot x , fy , i y t −1 = 0.694. This result indicates the greater potential for dynamic optimization to expand production by optimally choosing investment relative to the static model that takes investment as given. By year, average dynamic efficiency ranges from a low of ∆ t x , fy , i y = 0.474 in 2004 to a high of ∆ t x , fy , i y = 0.747 in 1991. Average static efficiency ranges from a low of Dot x , fy , i y t −1 = 0.592 in 2004 to a high of Dot x , fy , i y t −1 = 0.762 in 1999.

(

)

(

(

(

( )

)

(

(

)

)

)

)

Dynamic Efficiency and Productivity 203 Table 5.1 Estimates of Dynamic and Static Efficiency Dynamic

Static

∆t ( x t , f y t , i y )

# on Frontier

Dot ( x t , f y t , iy t−1)

# on Frontier

1991

0.747

6

0.722

2

1992

0.601

4

0.701

3

1993

0.567

1

0.682

2

1994

0.535

1

0.700

3

1995

0.546

1

0.706

2

1996

0.513

1

0.678

2

1997

0.544

1

0.687

1

1998

0.593

1

0.751

2

1999

0.547

1

0.762

2

2000

0.526

1

0.716

2

2001

0.478

1

0.630

1

2002

0.502

1

0.627

2

2003

0.486

1

0.632

2

2004

0.474

1

0.592

2

2005

0.492

1

0.666

2

2006

0.512

1

0.725

3

2007

0.517

1

0.748

3

2008

0.680

3

0.755

4

2009

0.532

1

0.727

3

2010

0.564

1

0.723

4

2011

0.569

1

0.651

2

Sub-periods 1991–1996

0.585

0.698

1997–2001

0.538

0.709

2002–2006

0.493

0.648

2007–2011

0.572

0.721

All years

0.549

0.694

The same number or more countries produce on the static frontier than produce on the dynamic frontier in every year except 1991 and 1992. This finding suggests that misspecified dynamics might be a source of bias in the estimation of technical efficiency, a finding consistent with Ahn and Sickles (2000), who found that technical efficiency follows an autoregressive process. Although not reported, Norway is on the dynamic

204 Färe, Grosskopf, Margaritis, and Weber frontier in 18 out of 21 years, and it produces on the static frontier in 20 out of 21 years. Table 5.1 also reports estimates of dynamic and static efficiency for four sub-periods. Dynamic efficiency is greatest during 1991–1996, which coincided with a global downturn, and lowest during the 2002–2006 expansion in the global economy. Static efficiency exhibits less variation between sub-periods and is highest during 2007–2011 and lowest in 2002–2006. Table 5.2 reports the geometric means of the dynamic and static productivity indices and their components of efficiency change and technical change for each year and for four sub-periods. For the dynamic model, productivity growth is positive in 14 out of 20 years and averages 1.7% over all years. The static model gives positive Table 5.2 Estimates of Dynamic and Static Productivity Change (Geometric Means) Dynamic

Static

Year

Malm

EFFCH

TECH

Malm

EFFCH

TECH

1991–1992

0.998

0.777

1.285

1.083

0.963

1.125

1992–1993

0.992

0.954

1.040

1.037

0.978

1.061

1993–1994

1.015

0.934

1.087

1.023

1.030

0.993

1994–1995

1.032

1.025

1.007

1.018

1.010

1.008

1995–1996

1.011

0.936

1.080

0.966

0.957

1.010

1996–1997

1.029

1.060

0.970

1.015

1.012

1.003

1997–1998

1.017

1.089

0.934

0.987

1.097

0.900

1998–1999

1.040

0.921

1.129

1.027

1.019

1.008

1999–2000

1.055

0.962

1.096

1.080

0.934

1.156

2000–2001

1.020

0.909

1.122

0.964

0.880

1.095

2001–2002

1.020

1.050

0.971

1.039

0.995

1.045

2002–2003

0.994

0.966

1.029

1.010

1.006

1.004

2003–2004

1.006

0.977

1.029

0.997

0.931

1.071

2004–2005

1.026

1.043

0.984

0.965

1.131

0.853

2005–2006

0.967

1.040

0.930

0.929

1.085

0.856

2006–2007

0.997

1.010

0.987

0.951

1.033

0.921

2007–2008

0.979

1.333

0.735

0.939

1.012

0.928

2008–2009

1.097

0.780

1.407

1.016

0.970

1.047

2009–2010

1.015

1.062

0.955

1.260

0.996

1.266

2010–2011

1.031

1.011

1.020

0.949

0.900

1.055

1991–1996

1.010

0.921

1.096

1.025

0.987

1.038

1997–2001

1.032

0.986

1.047

1.014

0.986

1.028

2002–2006

1.003

1.015

0.988

0.987

1.027

0.961

2007–2011

1.023

1.024

0.999

1.016

0.981

1.036

1991–2011

1.017

0.986

1.031

1.011

0.995

1.016

Dynamic Efficiency and Productivity 205 productivity growth in 10 out of 20 years and averages 1.1%. Increases in efficiency occur in 10 years for both the dynamic and the static models, but during the entire period there is a slight decline in efficiency, averaging 1.4% per year for the dynamic model and 0.5% for the static model. Positive technical progress occurs in 12 years for the dynamic model and 14 years for the static model. Average technical progress of 3.1% for the dynamic model and 1.6% for the static model more than offsets the declines in average efficiency. From Table 5.2 we see positive productivity growth in every sub-period for the dynamic model and in three out of four sub-periods for the static model. Productivity growth during the Great Recession period of 2007–2011 averages 2.3% for the dynamic model and 1.6% for the static model. During 2007–2010 efficiency increases by 2.4% for the dynamic model with 0.1% decline in technical progress. These results are consistent with evidence reported for the United States by Fernald (2015) and Petrosky-Nadeau (2013), who suggest that crises bring about productive resource reallocation. Foster et al. (2013) note that “[e]vidence shows this high pace of reallocation is closely linked to production dynamics.” We regard these results as illustrative; the empirical specification model is very parsimonious, we ignore discounting, and there are no connections of the intermediate product to the next period’s capital stock. A more sophisticated specification might result in larger differences between the comparative static and dynamic results. Nonetheless, these results illustrate what can be accomplished with dynamic productivity models, and these models can be expanded accordingly in several directions.

5.5. Conclusion In this chapter we have focused on at least some of the efforts to move measuring efficiency and productivity in a DEA framework from a static or comparative-static approach toward a more dynamic approach. The underlying links here are the distance functions, which have been the workhorses of the pioneering work motivated by CCD (1982). Since the distance functions represent technology, the first step is to specify technology in a dynamic framework that is still amenable to DEA-type estimation. A number of scholars have addressed this topic—a few of whom we have included here, with apologies to the many we did not. Our view of a dynamic technology is one in which decisions in the current (or past) period are explicitly linked and affect later periods. This includes notions of intermediate products, investment, time substitution, supply chain, and so on, and allows for possible reallocation across periods. This structure is also familiar from the network DEA literature. We believe the advantage of this approach is in better describing production processes and in providing more information to both analysts and policymakers. The resulting distance functions defined on these technologies we interpret as dynamic distance functions. We specify a many-period dynamic model in the spirit of

206 Färe, Grosskopf, Margaritis, and Weber Ramsey (1928), as well as an adjacent-period model familiar from the Malmquist productivity literature, and we provide an empirical illustration for the former. Extensions of the general setup are relatively straightforward for other distance function based productivity indices, both parametric and nonparametric. Although not explicitly included here, efficiency and productivity in the presence of good and bad outputs can also be cast in a dynamic setting; examples are cited in the introduction to the chapter. In the joint production framework, directional distance functions and Luenberger indices are appropriate building blocks. Future directions include the many refinements that have been developed in the comparative-static framework: statistical inference and possible bootstrapping, treatment of environmental variables, decompositions, and windows-type models, among many others.

Notes 1. CCD assume translog distance functions with identical second-order coefficients and input and output prices reflecting competitive outcomes. 2. This framework may also be useful in accommodating longer lags. For example, strong demand for energy and minerals from China and India in the first decade of the 2000s led Australian mining companies to begin capital expansion. Eslake (2011) attributed the decline in Australian mining sector productivity to be the result of the long lead times necessary to bring the new capital into full production. 3. Research by Pearl (2009) provides an important graphical exposition of how causal effects can be unraveled. 4. See also Kao (2013, 2014) for overviews of dynamic DEA and network DEA. We view dynamic DEA as a subset of network DEA. 5. For a survey of these conditions, see Färe and Primont (1995). Although we focus on measuring dynamic productivity using output sets, one can also measure dynamic productivity using input sets with the same regularity conditions. 6. This figure is from Färe and Grosskopf (1996). 7. This model is due to von Neumann; see Karlin (1959, 340). 8. We use the terms exogenous and endogenous in a nonstatistical sense. By exogenous we mean that those inputs are not variables, but rather data. By endogenous we mean that those factors are variables for which we solve. 9. This model is not consistent with the simple definition of productivity and productivity change as ratios of output and input quantity indices; see the discussion in O’Donnell (2012a, 2012b). 10. In a recent paper Diewert and Fox (2014) provide theoretical justification for why the geometric mean should be used in the Bjurek productivity index. Similar arguments apply to the Malmquist index. 11. Again, our use of exogenous/endogenous in this context is not statistical. Exogenous means given by observed data, and endogenous means solved for as part of the optimization problem. 12. The preceding problem is essentially a fixed base problem, which allows for transitivity but is base dependent; see Balk and Althin (1996) and Althin (2001). 13. Data and the GAMS programs used to estimate the static and dynamic models are available at http://cstl-hcb.semo.edu/wlweber.

Dynamic Efficiency and Productivity 207

References Ahn, S. C., and R. Sickles. 2000. “Estimation of Long-Run Inefficiency Levels: A Dynamic Frontier Approach.” Econometric Reviews 19: 461–492. Aigner, D. J., and S. F. Chu. 1968. “On Estimating the Industry Production Function.” American Economic Review 58: 226–239. Aiyar, S., C.-J. Dalgaard, and O. Moav. 2008. “Technological Progress and Regress in Pre-industrial Times.” Journal of Economic Growth 13(2): 125–144. Althin, R. 2001. “Measures of Productivity Changes: Two Malmquist Index Approaches.” Journal of Productivity Analysis 16: 108–128. Ang, F., and A. O. Lansink. 2014. “Dynamic Profit Inefficiencies: A DEA Application to Belgian Dairy Farms.” Bioeconomics Working Paper 2014/3, University of Leuven. http:// ageconsearch.umn.edu/bitstream/165693/2/BioeconWP_2014_3.pdf Balk, B. M. 1993. “Malmquist Productivity Indexes and Fisher Ideal Indexes: Comment.” Economic Journal 103: 680–682. Balk, B. M., and R. Althin. 1996. “A New Transitive Productivity Index.” Journal of Productivity Analysis 7: 19–27. Bjurek, H. 1996. “The Malmquist Total Factor Productivity Index.” Scandinavian Journal of Economics 98(2): 303–313. Bogetoft, P., R. Färe, S. Grosskopf, K. Hayes, and L. Taylor. 2009. “Dynamic Network DEA: An Illustration.” Journal of the Operational Research Society of Japan 52(2): 147–162. Caves, D., L. Christensen, and W. E. Diewert. 1982. “The Economic Theory of Index Numbers and the Measurement of Input, Output, and Productivity.” Econometrica 50(6): 1393–1414. Chen, C.-M., and J. Van Dalen. 2010. “Measuring Dynamic Efficiency: Theories and an Integrated Methodology.” European Journal of Operational Research 203: 749–760. Diewert, W. E. 1976. “Exact and Superlative Index Numbers.” Journal of Econometrics 4: 115–145. Diewert, W. E. 1992. “Fisher Ideal Output, Input and Productivity Indexes Revisited.” Journal of Productivity Analysis 3(3): 211–248. Diewert, W. E., and K. J. Fox. 2014. “Decomposing Bjurek Productivity Indexes into Explanatory Factors.” UNSW Australia Business School Research Paper No. 2014-33. Emrouznejad, A., and E. Thanassoulis. 2005. “A Mathematical Model for Dynamic Efficiency Using Data Envelopment Analysis.” Applied Mathematics and Computation 160(2): 363–378. Eslake, S. 2011. “Productivity: The Lost Decade.” Paper presented to the Reserve Bank of Australia Conference: The Australian Economy in the 2000s, Sydney, August. http://www. rba.gov.au/publications/confs/2011/pdf/eslake.pdf. Färe, R, and S. Grosskopf. 1992. “Malmquist Productivity Indexes and Fisher Ideal Indexes.” The Economic Journal 102(4): 158–160. Färe, R., and S. Grosskopf. 1996. Intertemporal Production Frontiers: With Dynamic DEA. Boston: Kluwer Academic. Färe, R., S. Grosskopf, B. Lindgren, and P. Roos. 1994. “Productivity Developments in Swedish Hospitals: A Malmquist Output Index Approach.” In Data Envelopment Analysis: Theory, Methodology and Applications, edited by A. Charnes, W. Cooper, A. Lewin, and L. Seiford, 253–272. Boston: Kluwer Academic. Färe, R., S. Grosskopf, and D. Margaritis. 2010. “Time Substitution with Application to DEA.” Journal of the Operational Research Society 62(7): 1420–1422.

208 Färe, Grosskopf, Margaritis, and Weber Färe, R., S. Grosskopf, D. Margaritis, and W. Weber. 2012. “Technological Change and Timing Reductions in Greenhouse Gas Emissions.” Journal of Productivity Analysis 37(3): 205–216. Färe, R., S. Grosskopf, D. Margaritis, and W. Weber. 2015. “The EU Stability and Growth Pact.” In Advances in Data Envelopment Analysis, edited by R. Färe, S. Grosskopf, and D. Margaritis, 75–86. Singapore: World Scientific-Now Publishers Series in Business. Färe, R., S. Grosskopf, and P. Roos. 1998. “Malmquist Productivity Indexes: A Survey of Theory and Practice.” In Index Numbers: Essays in Honour of Sten Malmquist, edited by R. Färe, S. Grosskopf, and R. R. Russell, 127–190. Boston: Kluwer Academic. Färe, R., and D. Primont. 1995. Multi-Output Production and Duality: Theory and Applications. Boston: Kluwer Academic. Färe, R., and G. Whitaker. 1996. “Dynamic Measurement of Efficiency: An Application to Western Public Grazing.” In Intertemporal Production Frontiers: With Dynamic DEA, edited by R. Färe and S. Grosskopf, 168–186. Boston: Kluwer Academic. Fernald, J. G. 2015. “Productivity and Potential Output before, during and after the Great Recession.” NBER Macroeconomics Annual 29(1): 1–51. Foster, L., C. Grim, and J. Haltiwanger. 2013. “Reallocation in the Great Recession: Cleansing or Not?” NBER Working Paper 20427. http://www.nber.org/papers/w20427. Fukuyama, H., and W. L. Weber. 2015a. “Nonperforming Loans in the Bank Production Technology.” In Quantitative Financial Risk Management: Theory and Practice, edited by Constantin Zopounidis and Emilios Galariotis, 46–70. Hoboken, NJ: John Wiley & Sons. Fukuyama, H., and W. L. Weber. 2015b. “Measuring Japanese Bank Performance: A Dynamic Network DEA Approach.” Journal of Productivity Analysis 44(3): 249–264. Fukuyama, H., and W. L. Weber. 2017. “Measuring Bank Performance with a Dynamic Network Luenberger Indicator.” Annals of Operations Research 250(1): 85–104. Fukuyama, H., W. L. Weber, and Y. Xia. 2016. “Time Substitution and Network Effects with an Application to Nanobiotechnology Policy for US Universities.” Omega 60: 34–54. Grifell-Tatje, E., and C. A. K. Lovell. 1995. “A Note on the Malmquist Productivity Index.” Economics Letters 47: 169–175. Kao, C. 2013. “Dynamic Data Envelopment Analysis: A Relational Analysis.” European Journal of Operational Research 227(2): 325–330. Kao, C. 2014. “Network Data Envelopment Analysis: A Review.” European Journal of Operational Research 239(1): 1–16. Kapelko, M., A. O. Lansink, and S. Stefanou. 2014. “Assessing Dynamic Inefficiency of the Spanish Construction Sector Pre-and Post-Financial Crisis.” European Journal of Operational Research 237: 349–357. Karlin, S. 1959. Mathematical Methods and Theory of Games, Programming and Economics. Reading, MA: Addison Wesley. Kemeny, J. G., O. Morgenstern, and G. L. Thompson. 1956. “A Generalization of the Von Neumann Model of an Expanding Economy.” Econometrica 24: 115–135. Koopmans, T. C. 1951. Activity Analysis of Production and Allocation. Cowles Commission Monograph No. 13. New York: John Wiley & Sons. Kumar, S., and R. R. Russell. 2002. “Technical Change, Technological Catch-Up, and Capital Deepening: Relative Contributions to Growth and Convergence.” American Economic Review 923(3): 527–548. Lansink, A. O., S. Stefanou, and T. Serra. 2015. “Primal and Dual Dynamic Luenberger Productivity Indicators.” European Journal of Operational Research 241(1): 555–563.

Dynamic Efficiency and Productivity 209 Nemoto, J., and M. Goto. 2003. “Measurement of Dynamic Efficiency in Production: An Application of Data Envelopment Analysis to Japanese Electric Utilities.” Journal of Productivity Analysis 19: 191–210. Nishimizu, M., and J. M. Page. 1982. “Total Factor Productivity Growth, Technological Progress and Technical Efficiency Change: Dimensions of Productivity Change in Yugoslavia 1965– 78.” Economic Journal 92: 920–936. O’Donnell, C. J. (2012a). “Aggregate Quantity Framework for Measuring and Decomposing Productivity Change.” Journal of Productivity Analysis 38: 255–272. O’Donnell, C. J. 2012b. “Nonparametric Estimation of Productivity and Profitability Change in U.S. Agriculture.” American Journal of Agricultural Economics 94(4): 873–890. O’Donnell, C. J. 2014. “Econometric Estimation of Distance Functions and Associated Measures of Productivity and Efficiency Change.” Journal of Productivity Analysis 41: 187–200. Pastor, J. T., and C. A. K. Lovell. 2005. “A Global Malmquist Productivity Index.” Economics Letters 88(2): 266–271. Pearl, J. 2009. Causality: Models, Reasoning, and Inference. Cambridge: Cambridge University Press. Petroskey-Nadeau, N. 2013. “TFP during a Credit Crunch.” Journal of Economic Theory 148: 1150–1178. Ramsey, F. P. 1928). “A Mathematical Theory of Saving.” Economic Journal 38: 543–559. Rungsuriyawiboon, S., and S. E. Stefanou. 2007. “Dynamic Efficiency Estimation: An Application to U.S. Electric Utilities.” Journal of Business and Economic Statistics 25(2): 226–238. Serra, T., O. Lansink, and S. E. Stefanou. 2011. “Measurement of Dynamic Efficiency: A Directional Distance Function Parametric Approach.” American Journal of Agricultural Economics 93(3): 756–767. Silva, E., and S. E. Stefanou. 2003. “Nonparametric Dynamic Production Analysis and the Theory of Cost.” Journal of Productivity Analysis 19(1): 5–32. Silva, E., and S. E. Stefanou. 2007. “Dynamic Efficiency Measurement: Theory and Application.” American Journal of Agricultural Economics 89(2): 398–419. Shephard, R. W. 1970. Theory of Cost and Production Functions. Princeton, NJ: Princeton University Press. Shephard, R. W., and R. Färe. 1980. Dynamic Theory of Production Correspondences. Cambridge, MA: Oelgeschlager, Gunn and Hain. Skevas, T., A. O. Lansink, and S. E. Stefanou. 2012. “Measuring Technical Efficiency in the Presence of Pesticide Spillovers and Production Uncertainty: The Case of Dutch Arable Farms.” European Journal of Operational Research 223: 550–559. Sueyoshi, T., and K. Sekitani. 2005. “Returns to Scale in Dynamic DEA.” European Journal of Operational Research 161: 536–544. Zucker, L. G., M. R. Darby, and J. S. Armstrong. 2002. “Commercializing Knowledge: University Science, Knowledge Capture, and Firm Performance in Biotechnology.” Management Science 48(1): 138–153.

Chapter 6

Produ ct i v i t y M easu rem ent i n Se c tors w it h Hard-t o -Me asu re Ou tpu t Kim Zieschang

6.1. Introduction The key challenge in measuring productivity is its inherently residual nature: productivity is output growth—including changes in output quality—over and above growth in intermediate and capital service inputs—including changes in input quality—while adjusting change in inputs for non-unitary returns to scale. Thus, using the national accounting term volume for the combined measure of change in quantity and quality, a key objective in compiling estimates of growth in output and intermediate consumption is to distinguish volume change from price change within the change in economic value aggregates, key among which is gross domestic product (GDP). Productivity thus purportedly measures the contribution of everything but output and input volume to evolution in the scale-adjusted output over input ratio. Everything else comprises, for example, technology, process, and environmental factors that are not embodied in the measure of output and input quality change. However, given a chosen scope of output, everything else also includes omitted inputs and measurement errors in output and input quantity and quality. This chapter is concerned with sectors having hard-to-measure output, but in measuring productivity in sectors with hard-to-measure output, we also are implicitly dealing with hard-to-measure intermediate input to the extent that those outputs are used in further production, changes in the quality of quasi fixed (capital) inputs, and correctly setting the boundary between what is an indicator of output and input quality and what is an indicator of evolution in technology, process, and environment.

212 Kim Zieschang The central heuristic of the theory and practice of productivity and efficiency measurement is an agricultural or manufacturing enterprise producing goods (output) with goods sourced from other producers (intermediate consumption), as well as labor and capital (primary services). Although produced (as opposed to primary) services have been part of output for the national accounts since Stone (1947), measuring the output and intermediate consumption of produced services has, as a rule, been challenged by problems with defining the volume, and even the nominal value of service outputs and inputs. As well, there have been long economics literatures on measuring both labor and (nonhuman) capital primary services, using the empirical capital services and capital accumulation paradigm defined in Jorgenson (1963). Practically speaking, quantity is in the units of the buyer-seller contract. Goods examples are a liter of fuel, a metric ton of grain, a car, a lathe, a building. Legal and other professional services are charged by the hour or by the job, while the principal quantity metric of financial services is the account, augmented in some instances by indicators of account servicing activity, such as check clearing. The quantity variable’s variations are directly reflected in a contractual pricing formula—the simplest of which being price times quantity—describing how the monetary value of the transaction between buyer and seller is determined. Quantity metrics are necessarily associated with conditioning metrics, or characteristics, describing what constitutes each transacted unit and that collectively comprise “product quality.” For example, fuel has an octane dimension; grain has species and grading dimensions; a car has quite a large number of dimensions such as size, power, interior and exterior finishes, and reliability ratings; a real estate asset is defined by building and lot size, location, quality of finishes, and local amenities/services. The value of legal services is conditioned by the reputation of the service provider in successful litigation. Financial services comprise indicators of the provision of liquidity, asset management, and insurance. When we declare a good or service difficult to measure, we are almost always dealing with a lack of data on the characteristics metrics—quality—associated with the quantity metric, and in some cases are not clear on what information to look for in acquiring these data. When variations in conditioning metrics affect the price of a given primary quantity transaction, we characterize them as describing input, output, or process quality, thus connecting us with the economics literature on product quality measurement and its impact on volume transacted. We can identify the following hard-to-measure sectors for productivity measurement: • High technology industries • Real estate • Services, notably • Distributive services • Financial services • Banking • Insurance

Productivity Measurement in Sectors 213

• Health care • Education

A first class of issues among these hard-to-measure cases is where the characteristics metrics on outputs and inputs are changing (i.e., “quality” is changing) between periods or places, along with quantities, in a productivity comparison. Quality change is particularly evident in service activities, but goods undergoing rapid changes in technology, such as information and communications technology (ICT) equipment, are also among well-known hard-to-measure cases of this type. Bosworth and Triplett (2007) provide a good overview of the state of play in services measurement, while Byrne, Fernald, and Reinsdorf (2016) explore issues in measuring the output of high-technology industries, and across all industries, in measuring the capital input from the accumulation of ICT equipment and intangible assets such as intellectual property. A second element of “hardness to measure” is lack of sufficiently frequent transaction data to permit market valuation, and is particularly relevant for certain durables. A key example of this is real estate, a combination of land, improvements to land, and structures with location-specific, time-varying quality dimensions whose market prices are determined by sale transactions separated by years or decades. A third “hardness to measure” issue is lack of information on the production and accumulation of intellectual property assets, which are relatively recent additions to national accounting standards aimed at capturing at least some of Romer’s (1990) and others’ “endogenous technological change.”1 These types of products may be both output and intermediate consumption in the calculation of GDP, but as underscored in Byrne et al. (2016), our overview of hard-to-measure sectors would be incomplete without also considering the human and nonhuman capital measurement problems within the denominator, input component of productivity measures.

6.2. Analytics of Hard-to-Measure Goods, Services, and Assets We begin with the economic analytics of accounting for changes in quality and environmental characteristics in measuring productivity.

6.2.1. Primal Technology: Distance Functions Distance functions are representations of production functions expressly designed to measure efficiency and productivity by looking at how much less input than we observe would be required to generate observed output, or how much more than the output we

214 Kim Zieschang observe could be produced with observed inputs. We begin by defining the following variables: y the vector of output quantities x the vector of input quantities provided by other producers (intermediate inputs) K = K H , K N ′ the vector of human (H) and nonhuman (N) capital stocks yielding primary input services the vector of characteristics metrics (“quality”) for outputs γ > 0 χ > 0 the vector of characteristics metrics (“quality”) for intermediate inputs κ = κ H , κ N ′ > 0 the vector of characteristics metrics (“quality”) for capital (human and nonhuman asset) stocks ϑ > 0 the vector of conditioning metrics (“characteristics” or “quality”) for technology and environment. The quality and environment variables are presumed to be defined, with little or no sacrifice of generality, so that ( γ , χ, κ , ϑ ) > 0. We define the primal technology as the set T of feasible ( y , x , K , γ , χ, κ , ϑ ). In the interest of focusing on key points, we make the following free disposal (monotonicity) assumptions concerning technology T:

(

) ( , χ , κ ) ≥ ( x , K , χ, κ ) then ( y , x , K

)

If ( y , x , K , γ , χ, κ , ϑ ) ∈T and y 0 , γ 0 ≤ ( y , γ ) then y 0 , x , K , γ 0 , χ, κ , ϑ ∈T ;

(

If ( x , K , χ, κ ) ∈T and x 0 , K 0

0

0

0

0

)

, γ , χ0 , κ 0 , ϑ ∈T .

Thus armed, we can choose the economist’s production function from an array of functional metrics on T. These functional metrics are called distance functions, and for any given technology T can be set up in various configurations, depending on the analytical focus. We will introduce three such metrics briefly—the input distance function, the output distance function, and the capacity utilization function—before narrowing detailed discussion to the third option in the remainder of the chapter. The input distance function is defined2,3

{

}

dx ( y , x , K , γ , χ, κ , ϑ ) ≡ min λ : ( y , λ x , K , γ , χ, κ , ϑ ) ∈T .

( y , x, K , γ , χ, κ , ϑ ) is feasible (that is, ( y , x, K , γ , χ, κ , ϑ ) ∈T ) if and only if dx ( y , x , K , γ , χ, κ , ϑ ) ≥ 1. By definition dx is positively linear homogeneous in x: dx ( y , λ x , K , γ , χ, κ , ϑ ) = λdx ( y , x , K , γ , χ, κ , ϑ ) for λ > 0. Given the free disposal assumptions on T, dx is nonincreasing in ( y, γ ) and nondecreasing in ( x , K , χ, κ ). The output distance function is defined as

{

}

d y ( y , x , K , γ , χ, κ , ϑ ) ≡ min δ : ( y / δ, x , K , γ , χ, κ , ϑ ) ∈T .

( y , x, K , γ , χ, κ , ϑ ) is feasible if and only if d y ( y , x, K , γ , χ, κ , ϑ ) ≤ 1. By definition d y is positively linear homogeneous in y: d y ( λ y , x , K , γ , χ, κ , ϑ ) = λd y ( y , x , K , γ , χ, κ , ϑ )

Productivity Measurement in Sectors 215 for λ > 0. Given the free disposal assumptions on T, d y is nonincreasing in ( x , K , χ, κ ), nondecreasing in ( y, γ ). The capacity utilization function is defined as

{

}

d yx ( y , x , K , γ , χ.κ , ϑ ) ≡ min η : ( y / η, x / η, K , γ , χ, κ , ϑ ) ∈T .4

d yx is positively linear homogeneous in ( y , x ): d yx ( λ y , λ x , K , γ , χ, κ , ϑ ) = λd yx ( y , x , K , γ , χ, κ , ϑ ) for η > 0. If ( y , x , K , γ , χ.κ , ϑ ) is feasible, then d yx ( y , x , K , γ , χ, κ , ϑ ) ≤ 1.5 Given the free disposal assumptions on T, d yx is nondecreasing in ( y, γ ) and nonincreasing in ( x , K , χ, κ ). Notwithstanding our capacity utilization function moniker for the third distance function, all three production functions have a capacity utilization interpretation. Full capacity is indicated when the associated distance function is equal to 1 and the degree of less than full capacity is indicated by a distance function value less than 1. For the input distance function, capacity is defined by a set of outputs y conditional −1 on ( K, γ , χ, κ , ϑ ). dx ( y , x , K , γ , χ.κ , ϑ ) (the reciprocal of the input distance function) measures the utilization of inputs in producing y given ( K, γ , χ.κ , ϑ ), where full capacity −1 (efficiency) is given by dx ( y , x , K , γ , χ.κ , ϑ ) = 1 and the degree of less than full capacity −1 is given by dx ( y , x , K , γ , χ.κ , ϑ ) < 1. For the output distance function, capacity is defined by the degree to which production of outputs y can be feasibly scaled up conditional on ( x , K , γ , χ.κ , ϑ ) (i.e., the degree to which they lie below the output frontier), where d y ( y , x , K , γ , χ.κ , ϑ ) = 1 indicates full capacity and d y ( y , x , K , γ , χ.κ , ϑ ) < 1 indicates the degree of less than full capacity. For the capacity utilization function, capacity indicates the degree to which outputs and intermediate inputs ( y , x ) can be feasibly scaled up conditional on capital stocks and characteristics ( K, γ , χ.κ , ϑ ). Our capacity utilization terminology for this distance function follows the traditional notion that capacity is set by quasi-fixed inputs, namely human and nonhuman capital stocks K = K H , K N ′ , which are the sources of primary services in the national accounts, and whose compensation is value added, the sum of which over all producers in the economy is GDP from the production approach.6 d yx ( y , x , K , γ , χ, κ , ϑ ) = 1 thus indicates full capacity—that value added has reached what is commonly known as potential GDP—and d yx ( y , x , K , γ , χ, κ , ϑ ) < 1 indicates the degree of less than full capacity (capital stock underutilization and GDP below potential).

6.2.2. Quality Adjusted Productivity Applying the three production metrics to comparisons of two situations, say, y 0 , x 0 , K 0 , γ 0 , χ0 , κ 0 , ϑ 0 and y1 , x1 , K 1 , γ 1 , χ1 , κ1 , ϑ1 , where 0 and 1 might reference distinct time periods or geographical regions or industrial activities or enterprises, produces three variants of multifactor productivity when comparing successive points on the production possibilities frontier of T. The input and output distance function approaches have been extensively applied in the efficiency and productivity literature, mostly for individual enterprises within industries, but we focus here only on

(

)

(

)

216 Kim Zieschang productivity based on the third, the capacity utilization function, because of its relationship to value added and GDP. We therefore consider the following multifactor productivity index M

01 s y,x

=

( ) (y , x , K , γ , χ ,κ , ϑ )

d 0yx y s , x s , K s , γ s , χ s , κ s , ϑ 0 d

1 yx

s

s

s

s

s

s

1

where s indicates information from a reference situation (e.g., time period, locality, or 01 s type of activity/industry). M y , x measures the change in the capacity frontier for y s , x s as environmental characteristics and perhaps the functional form of technology change from ϑ 0 to ϑ1 given capital stocks K s and the characteristics of outputs, intermediate 01 s inputs, and capital stocks γ s , χ s , κ s . M y , x throws the residual nature of productivity measurement into relief, as the measure of that part of the change in capacity utilization 01 s not arising from the change in output or input volume. M y , x is interpreted as follows:

(

(

)

)

• If the capacity utilization function declines in moving from technology 0 to tech01 s nology 1 (M y , x > 1), so that a given vector of outputs, intermediate consumptions, capital stocks, and the associated characteristics are less efficiently produced under technology 1, then technology 1 must be more productive than technology 0 (given K , there is greater capacity to expand ( y , x ) under technology 1 than under technology 0). • If the capacity utilization function increases in moving from technology 0 to tech01 s nology 1 (M y , x > 1), so that a given vector of outputs, intermediate consumptions, and capital stocks, and the associated characteristics are more efficiently produced under technology 1, then technology 1 must be less productive than technology 0 (given K , there is less capacity to expand ( y , x ) under technology 1 than under technology 0). When the reference situation s is chosen as s = 0, we say the productivity comparison has a Laspeyres perspective, inspired by the weighted linear and still widely used Laspeyres (1871) index number. When s = 1, we say the productivity comparison has a Paasche perspective, inspired by the weighted harmonic and still widely used Paasche (1874) index number. If we take the geometric mean of Laspeyres and Paasche perspective multifactor indices, we will term this a Fisher perspective comparison, inspired by Fisher’s (1922) ideal index number, which is the geometric mean of Laspeyres and Paasche index numbers. With these preliminaries, we can write the Fisher- perspective multifactor productivity index

( (

) )

( (

d 0yx y 0 , x 0 , K 0 , γ 0 , χ0 , κ 0 , ϑ 0 d 0yx y1 , x1 , K 1 , γ 1 , χ1 , κ1 , ϑ 0 01 01 M y,x = 1 0 0 0 0 0 0 1 × 1 1 1 1 1 1 1 1 d yx y , x , K , γ , χ , κ , ϑ d yx y , x , K , γ , χ , κ , ϑ

) )

1

2 . (6.1)

Aside from symmetry in binary comparisons (not having to choose between the Laspeyres and Paasche perspectives in a comparison by incorporating both), if there

Productivity Measurement in Sectors 217

(

) ( (

)

were no change in the quality characteristics γ 0 , χ0 , κ 0 = γ 1 , χ1 , κ1 , and under the s assumption that we only observe efficient states where d yx y s , x s , K s , γ s , χ s , κ s , ϑ s = 1,7 the Fisher perspective productivity index also is associated with empirically powerful and useful exact index number results. These results allow us to calculate multifactor productivity using a formula depending only on the prices and quantities for outputs, intermediate consumptions, and capital stocks y s , x s , K s , by applying the Translog identity of Caves, Christensen, and Diewert (CCD) (1982). The CCD (1982) result in our capacity utilization function context is the following. Assuming

(

)

)

• there is no quality change in outputs, intermediate inputs, or capital stocks so that γ 0 , χ0 , κ 0 = γ 1 , χ1 , κ1 ; • producers maximize value added, V ≡ p ′y − v ′x given ( K, γ , χ, κ , ϑ ), where p is the given vector of basic prices8 of outputs and v is the given vector of purchasers’ prices9 of intermediate inputs; • producers minimize primary input cost, C ≡ u ′K given ( y , x , γ , χ, κ , ϑ ), where u = ρι + δ is the given price vector of quasi-fixed inputs (capital stocks), with ρ the enterprise cost of capital, ι is a vector of ones of same dimension as K , and δ is the vector of depreciation rates for capital stocks K ; and • for each period, location, or activity s the capacity utilization function s d yx y , x , K , γ , χ, κ , ϑ has a Translog flexible functional form (the log of the capacity utilization function is quadratic in the logs of its arguments), where the coefficients second-order terms of the Translog functions do not vary from time to time;

(

) (

(

)

)

the following Törnqvist multifactor productivity index number is exact 1 01 01 M y , x T = exp w 0y + w1y ′ ln y1 − ln y 0 2 1 0 1 1 ′ × exp − w x + w x ln x − ln x 0 2 1 0 ′ × exp − ε d yx , K w K0 + ε1d yx ,K w1K ln K 1 − ln K 0 2

(

)(

(

)

)(

)

(

)(

(6.2)

)

where (a) if is the Hadamard or elementwise product operator on two vectors of the same dimension,10

( x /(p

) x )

w sy = p s y s / p s ′ y s − v s ′ x s s x s K

s

s

s′

s

w =v y −v s s s′ w = uK K /uK K s

s′

s

218 Kim Zieschang are the vectors of, respectively, the value-added shares of outputs, the value-added shares of intermediate inputs, and the primary input cost shares of capital services; and,

ε ds yx , K = −

(b)

s K s ′ ∇K d yx s s y s ′ ∇ y d yx + x s ′ ∇ x d yx

(

)

is the elasticity of outputs and intermediate consumptions y s , x s with respect to the scale of quasi-fixed inputs (capital stocks) K s conditional on the quality γ s , χ s , κ s , ϑ s of output y s , intermediate consumption x s, capital stock κ s , and technology/environment ϑ s . The nonparametric expression for the elasticity of scale ε ds y ,x , K derives as follows. Suppose λ is a scaling factor for capital stocks K s and θ is a scaling factor for outputs and intermediate consumptions y s , x s such that

(

(

)

)

(

)

s d yx θ y s , θx s , λK s , γ s , χ s , κ s , ϑ s = 1.

We know from the maintained assumptions that value added is maximized and primary input cost is minimized that, respectively, θ = 1 and λ = 1. Take the total differential of the preceding expression with respect to θ and λ and solve for ε ds y ,x , K = dθ / dλ at λ = 1 as

ε

s d yx ,K

s K s ′ ∇K d yx dθ = = − s′ . s s dλ y ∇ y d yx + x s ′ ∇ x d yx

If, in addition, the enterprise is subject to eventual decreasing returns to scale11 and also maximizing “super surplus” p s ′ y s − v s ′ x s − u s ′ K s ,12 the CCD (1982) results can be used to determine the elasticity of scale as

ε ds yx , K =

under the assumption that we observe

(

us′ K s ps′ y s − v s′ x s us′ K s

p ′ y s − v s′ x s s

≤ 1.13

)

However, if the characteristics γ s , χ s , κ s change for s = 0, 1, as they almost certainly will, to account for their impact on productivity we must know the gradients of d y , x with respect to γ s , χ s , κ s to apply the CCD (1982) Translog identity .

(

)

We can obtain these gradients by estimating an assumed parametric form for d y , x and taking the characteristics or quality gradients of the estimated capacity utilization function. Alternatively, we can proceed to a “semi-nonparametric” result generalizing the CCD (1982) translog identity. If γ s , χ s , κ s reliably affect the prices the producer receives for y s , x s , K s within each period/situation 0,1, then using a result from Fixler and Zieschang (1992):14

(

)

(

)

Productivity Measurement in Sectors 219 Under the same assumptions as the CCD (1982) translog identity, plus additionally that • there are (presumed known) hedonic price functions for, respectively, outputs, intermediate inputs, and capital service inputs as p s = Pys γ s , v s = Pxs χ s , and u s = PKs (κ s )

( )

( )

the following Törnqvist quality adjusted multifactor productivity index number is exact 1 01 01 M y , x T = exp w 0y + w1y ′ ln y1 − ln y 0 2 1 0 × exp − w x + w1x ′ ln x1 − ln x 0 2 1 0 ′ × exp − ε d yx , K w K0 + ε1d yx , K w1K ln K 1 − ln K 0 2 1 × exp ω 0γ + ω1γ ′ ln γ 1 − ln γ 0 2 1 0 1 ′ 1 × exp − ω χ + ω χ ln χ − ln χ0 2

(

)(

(

)

)(

)

(

)(

(

)(

(

)

)

)(

)

(

)(

1 ′ × exp − ε 0d yx , K ω 0κ + ε1d yx , K ω1κ ln κ1 − ln κ 0 (6.3) 2

)

where

( ) ( ) (χ )′ x / ( p ′ y − v ′ x ) (κ )′ K / u ′ K .

ω sy = ∇ γ s Pys γ s ′ y s / p s ′ y s − v s ′ x s

ω sx = ∇χs Pxs

ω sK = ∇κ s PKs

s

s

s

s

s

s

s

s

s

s

So, assuming we only observe efficient states, always on the production frontier, then aside from needing to know the characteristics gradients of the hedonic functions p s = Pys γ s , v s = Pxs χ s , and u s = PKs κ s , the capacity utilization productivity index can be written as a nonparametric index number when enterprises in the economy are presumed to be maximizing super surplus. In fact, the quality varying multifactor productivity index 01 01 M y , x Τ of equation (6.3) is the constant quality Törnqvist multifactor productivity index

( )

( )

( )

01 01

M y , x of equation (6.2) multiplied by the output-input quality adjustment factor

1 ′ 01 01 Q y , x T ≡ exp ω 0γ + ω1γ ln γ 1 − ln γ 0 2 1 0 ′ × exp − ω χ + ω1χ ln χ1 − ln χ0 2

(

)(

(

)

)(

)

1 0 ′ × exp − ε d yx , K ω 0κ + ε1d yx , K ω1κ ln κ1 − ln κ 0 2

(

)(

) .

220 Kim Zieschang 01 01

It also is evident from this that Q y , x Τ decomposes into three quality adjustment factors: 1 ′ • for output, exp ω 0γ + ω1γ ln γ 1 − ln γ 0 , 2 1 0 ′ • for intermediate consumption, exp − ω χ + ω1χ ln χ1 − ln χ0 , and 2 1 ′ • for capital stocks, exp − ε 0d yx , K ω 0κ + η1d yx , K ω1κ ln κ1 − ln κ 0 . 2

(

)(

)

(

(

)(

)

)(

)

Thus, quality improvements in output increase, and quality improvements in intermediate inputs and capital stocks decrease, multifactor productivity.

6.2.3. Relationship to Deflation Techniques Used in Productivity Accounting 01 01

Productivity accountants typically obtain the quantities in the productivity index M y , x T via deflation of current price monetary values from the national accounts. Thus, if the current price monetary flow for the ith output is E tyi = ptyi yit , for the jth intermediate t consumption is Exjt = pxjt x tj , and for the kth capital rental is EKk = ukt K kt , then the relative change in nominal flows also can be written as

ptyi 1 ′ = 0 exp − ω 0γ i + ω1γ i ln γ 1i − ln γ i0 0 E yi p yi 2 E tyi

)(

)

1 yt ′ × exp ω 0γ i + ω1γ i ln γ 1i − ln γ i0 i0 2 yi

(

(

)(

)

pxjt 1 0 1 ′ 1 0 = 0 exp − ω χ j + ω χ j ln χ j − ln χ j 0 2 Exj pxj Exjt

(

)(

)

t 1 xj ′ × exp ω 0χ j + ω1χ j ln χ1j − ln χ0j 0 2 xj

(

)(

)

(

t ukt EKk 1 0 0 1 1 = 0 u 0 exp − 2 ε d yx , K ω κk + ε d yx , K ω κk EKk k

(

)(

′

) (ln κ

1 k

)

− ln κ 0k

1 Kt ′ × exp ε 0d yx , K ω 0κ + ε1d yx , K ω1κ ln κ1 − ln κ 0 k0 . 2 Kk

)

Productivity Measurement in Sectors 221 By implication, the quality-adjusted quantities in the productivity index (6.3) can be obtained by substituting for them nominal expenditures deflated by quality adjusted price indices: 1 yt exp ω 0γ i + ω1γ i ′ ln γ 1i − ln γ i0 i0 2 yi t t E yi p yi 1 0 1 ′ 1 0 = 0 0 exp − ω γ i + ω γ i ln γ i − ln γ i 2 E yi p yi t 1 xj exp ω 0χ j + ω1χ j ′ ln χ1j − ln χ0j 0 2 xj t t Exj pxj 1 0 1 ′ 1 0 = 0 0 exp − ω χ j + ω χ j ln χ j − ln χ j Exj pxj 2 t 1 K ′ exp ε 0d yx , Kk ω 0Kk + ε 0d yx , Kk ω1Kk ln κ1k − ln κ 0k k0 2 Kk t t 1 0 EKk uk ′ 0 1 1 1 0 = 0 0 exp − ε d yx , K ω Kk + ε d yx , K ω Kk ln κ k − ln κ k . EKk uk 2

(

)(

)

(

(

)(

)

)(

)

)

(

(

)(

)( (

)

)(

)

6.2.4. Environment Variables ϑ In the semi-nonparametric results in the preceding, we do not include the environment 01 variables ϑ s in the productivity quality adjustment index number Q 01 because if they y,x T do not affect the hedonic price equations of the outputs, intermediate inputs, and/or capital stocks, they are subsumed into the change in technology between 0 and 1 and thus have no independent effect on the semi-nonparametric index. Of course, many hedonic evaluations assessing the economic value of environment-type variables find that these variables do affect the prices of market products.15 We will assume that if this is the case, such environmental variables are already included in the characteristics vectors s γ s , χ s , κ s . Thus ϑ is the vector of environmental and technology metrics whose variations do not affect the prices of outputs and inputs and thus are in this sense “disembodied” from those outputs and inputs.16 Failure to include environmental variables that affect the prices of market outputs and inputs in the hedonic equations describing the relationship between prices and the price-determining characteristics of those products constitutes a misspecification that will lead to bias in the residual productivity measure of the impacts of “disembodied” factors ϑ s .

(

)

6.2.5. Inefficiency A discussion point concerning the preceding result is that assuming value-added maximization conditional on primary inputs (capital stocks) and primary input cost

222 Kim Zieschang minimization conditional on output and intermediate consumption means that all observed data points (e.g., s = 0, 1) are conditionally efficient and lying on the boundary of the technology T, but that the enterprise is not necessarily operating at efficient scale. The fully nonparametric result hinges on the further requirement that super surplus be maximized, which in turn requires nonincreasing returns to the scale of capital stocks at the observed outputs, intermediate inputs, and capital stocks. If we further examine the derivation of the implicit rental rates of capital stocks u from the cash flow equation, we are further led to the conclusion that under the assumption of super surplus maximization and price taking we will only observe unitary, constant returns to scale. Thus in the CCD (1982) nonparametric result and the Fixler and Zieschang (1992) extension, inefficiency or super- efficiency is implied, not at the observed situations, where d 0y , x y 0 , x 0 , K 0 , γ 0 , χ0 , κ 0 , ϑ 0 = 1 and d1y , x y1 , x1 , K 1 , γ 1 , χ1 , κ1 , ϑ1 = 1, but only at the unobserved perturbations d1y , x y 0 , x 0 , K 0 , γ 0 , χ0 , κ 0 , ϑ1 ≤ 1 and d 0y , x y1 , x1 , K 1 , γ 1 , χ1 , κ1 , ϑ 0 ≤ 1. Efficiency, like multifactor productivity, is a fundamentally residual concept and thus is ultimately a measure of our ignorance, particularly regarding incomplete information on the characteristics variables γ s , χ s , κ s , ϑ s and capital stocks K s . For nonhuman capital stocks specifically, this pertains to high technology tangible assets and imperfectly measured intangibles, such as intellectual property17 and goodwill and marketing assets.18 We could therefore expect that the more comprehensively we measure the characteristics vectors, including ultimately the “disembodied” environmental characteristics, and include their effects in the productivity measure, the smaller the productivity and efficiency change we are likely to calculate. Thus, efficiency (like productivity) is not so much a concept in itself, but an indicator of the completeness of model specification.19

(

(

)

(

)

(

(

)

)

)

6.3. Empirical Issues in Hard-to-Measure Goods and Services 6.3.1. Parametric or Semi-Nonparametric? Econometric estimation of a parametric functional form for the primal capacity utilization function in the productivity index (6.1) is ultimately constrained by the size of the data sets available, particularly if the functional form is flexible. The context generally will be industry-level time series data at quarterly frequency. Given the large number of variables for which parameters are needed, degrees of freedom are likely to be exhausted quickly and the reliability of estimates thereby impaired. The semi-nonparametric index number results (6.3) for the Törnqvist index numbers presented here still require period by period hedonic estimates of some type, which usually (and least restrictively) are based on cross-sectional firm-or product-level data from given time periods, though the hedonic price equations can be estimated from

Productivity Measurement in Sectors 223 industry time series under restrictive assumptions. Size of data set may be significantly larger in the cross-sectional than time-series context (but also might not be large, as in concentrated industries), and the degree-of-freedom demands of hedonic equations potentially more parsimonious, as they are often less saturated with parameters than distance functions and their value-added and cost-function duals. Ultimately, it appears that better accounting for quality change in productivity indices will require more detailed enterprise-and/or establishment-level time series data— both time series and cross-sectional—from panel surveys.

6.3.2. Availability of Data on Key Variables The residual nature of productivity measures, particularly multifactor productivity indices, makes them sensitive to excluded intermediate and, particularly, capital inputs, as well as inaccurate adjustments for output quality change, because of lack of data on characteristics variables, or insufficient observations to identify a hedonic model for which price on the left-hand side is dependent on many characteristics variables on the right-hand side. Based on equation (6.3), within the Törnqvist measurement framework of this chapter, an omitted output characteristic for a given activity first drops the output factor corresponding to that quality characteristic from the index and, second, probably affects the estimated gradient(s) of the output price hedonic equation(s) Pys (γ s ) with respect to other characteristics, thus affecting the weights of the other characteristics in the index. If the characteristic shows rapidly improving quality of that output, its effect on the overall productivity picture will be attenuated by the degree to which the remaining included characteristics are correlated with it. If the index pertains to a particular industrial activity, the aggregate effect of its omission on productivity at the economy-wide level may be attenuated further to the extent that that output is used as an intermediate consumption by another activity. This is evident from the negative sign on the exponents in the productivity index’s quality-adjusted intermediate consumption factor:

1 1 exp − w x0 + w1x ′ ln x1 − ln x 0 × exp − ω 0χ + ω1χ ′ ln χ1 − ln χ0 2 2

(

)(

)

(

)(

) .

That is to say, an understatement of output and thus productivity in the given activity as a result of understatement in output quality change will result in overstatement of productivity in activities using that output as an intermediate consumption, attenuating the impact of the error on total economy productivity. An omitted capital stock input in the multifactor productivity index for a given activity, such as an accumulating intellectual property position, will drop the input factor for that capital stock and could affect the returns to scale adjustment for all inputs in that activity (depending on how the returns to scale ε ds yx , K is calculated). The aggregate effect on GDP is attenuated by the extent to which the output of that activity is itself

224 Kim Zieschang accumulated as a durable capital asset (e.g., research and development [R&D] services or mineral exploration services): an understatement of the accumulation of that durable capital asset will overstate productivity change in all activities using that asset. On the other hand, if the intellectual property output is not accumulated as a nonhuman asset but instead used as an intermediate input of other activities, its exclusion understates the value added of the intellectual output and overstates the value added of sectors using the intellectual product service. As a rule, the distinction between accumulating intellectual property assets and provision of intellectual services is decided by whether intellectual services such as R&D are generated for own use (capital formation final expenditure), sold to other producers (intermediate consumption), or sold to households as consumption (consumption final expenditure).

6.3.3. When Hedonic Quality-Adjusted Productivity Becomes a “Matched Models” Index At base, price and quantity index methodology conventionally relies on recording the prices and quantities of exactly the same collection of products over time, as in the CCD (1982)–type result of equation (6.2). This traditional approach is sometimes called a “matched models” method. Our superlative Törnqvist Index approach to multifactor, GDP productivity still captures change in the importance of available varieties of products through changes in the value-added and primary-input cost shares appearing in the exponents of equation (6.2). Index compilers then deal with cases where they lack observations on the quantities and prices of appearing or disappearing product varieties. This “lack of comparable varieties” problem seems to become less quantitatively important in a time series context as the frequency of data recording increases, because new varieties often enter and leave markets with very low market shares, depending on how powerful the price dynamics and change in market share are in the market penetration/exit period. This said, the frequency of price data recording often is limited by the cost of data acquisition, and the entering (exiting) share may not be captured soon enough (late enough). In non-time- series contexts, such as geographic (e.g., international or interregional) productivity comparisons the “lack of comparable varieties” problem remains unavoidable. The “lack of comparable varieties” issue is the “replacement” version of the “new and disappearing varieties” issue, about which more in a moment, where the introduction of a new variety or class of product with updated characteristics replaces a specific old variety or class. The framework of this chapter accommodates the lack of comparable varieties issue. Equation (6.3) reduces to equation (6.2) if the set of varieties or classes of products is unchanged over the comparison, but their quantities and prices change. If only some items change in quality, then only those items require a quality-adjusted treatment within the quality-adjusted productivity index of equation (6.3).

Productivity Measurement in Sectors 225

6.3.4. Truly New Products, Disappearing Products, and Changes in the Scope of Output, Intermediate Inputs, and Primary Services Equations (6.2) and (6.3) do not address changes in the scope of products—the number of varieties transacted. This is the purest form of the new and disappearing goods problem that does not imply replacement of a given existing product variety with a new one having different characteristics, but rather is the emergence of a new category of product or complete disappearance of an old category. The preferred conceptual treatment of a new or disappearing product is to impute its reservation price in the period prior to introduction or after disappearance. The reservation price is the lowest price at which just one unit of output would be produced, other things equal, or the highest price at which just one unit of intermediate or primary service input would be used, other things equal.20

6.3.5. Output and Intermediate Consumption Measurement Issues by Sector Byrne et al. (2016) provide an assessment of productivity measurement issues in the US economy over three successive eras: 1978–1995, 1995–2004, and 2004–2015. The US economy is large and diversified, with significant high technology and service sectors, and thus reasonably reflects the state of multifactor productivity measurement worldwide. Their findings, overall, suggest that better measurement of product quality, as well as of intellectual property and high technology assets, would have had a mildly salutary effect on productivity trends over the last 30 years, on the order of hundredths of a percentage point. However, these improvements would not have changed the historical profile or current trend in multifactor or labor productivity. In particular, the US productivity slowdown since 2004 remains intact after plausible adjustments are made to outputs and inputs for technical progress in a variety of industries, from high technology to mining to online retailing, social media, and the sharing economy.

6.3.5.1. High Technology Industries The high technology industries are principally the information and communications technology (ICT) group, comprising computers and peripherals; communications equipment; other information systems; and software. Quality change in computers and peripherals has been studied since the mid-to late 1980s, pioneered by the US Bureau of Economic Analysis (Cole et al. 1986; Dulberger 1989). Steady improvements have been made since, though over 2005–2014 (1) import substitution has been imperfectly captured in US data as domestic production has given way to nonresident producers, and (2) quality adjustment has been much less accurate

226 Kim Zieschang on imports than on domestic production. Byrne et al. (2016) used the computers and peripherals price index from communications equipment to deflate the nominal output of these goods.

6.3.5.2. Construction and Real Estate Assets Construction output is imminently tangible and visible, and seemingly easily measurable, but for two issues. The first is that data are expensive to collect because construction happens at a constantly changing set of geographical locations with projects at various stages of completion, and multiple personal visits are usually necessary to collect project data. The second is that constructions often build custom products that are not directly comparable with each other in a given time period or with other constructions completed in the past. Thus, the “matched models” approach to measuring price and volume change does not apply, and the hedonic approach to quality adjustment must contend with a large array of product characteristics that threaten to overwhelm the available sample of projects with the need to estimate parameter coefficients for large numbers of price determining characteristics. Sveikauskas et al. (2014) provide a comprehensive state of play for productivity measurement in construction. While there are active rental markets for property and structures capital in place, the related problem of measuring the price and volume of real estate assets in place is a still a developing field, which is partly susceptible to matched models methodology, but for staggered and infrequent property transactions that make the valuation of stocks and measurement of depreciation problematic. There is a recent set of international recommendations on measuring residential property prices (and thus via deflation the volume of the residential property stock).21 Graf and Silver (2014) provide an overview of thinking on measuring commercial property capital.

6.3.5.3. Services 6.3.5.3.1. Distributive Services Nominal distributive services are measured as the aggregate gross margins of retailers and wholesalers. Distributive service vendors sell marketing and display, transactions, and, more recently, logistics services to purchasers of goods as well as services. As a rule, the volume of distributive services is measured by the volume of goods and services these enterprises sell, which arguably captures aspects of the transactions and logistics services they produce. While goods and services transaction volume is perhaps not an adequate treatment of distributors’ marketing and display functions, aspects of these services may be captured by scope indicators such as number of product varieties marketed. This sector has undergone significant upheaval over the last 20 years with the advent and rapid increase in market share of online retailers and wholesalers, who display their wares on video screens rather than marketing physical product samples and fulfilling purchase transactions from inventory in brick and mortar stores and showrooms. Though not totally overwhelmed, the brick and mortar retail business

Productivity Measurement in Sectors 227 model has lost ground to web distribution by computerized logistics directing massive regional order-fulfillment operations and high-speed shipping modalities such as air freight and efficient local package delivery. Services have begun to enter the same web- enabled distributive modality through “gig economy” enterprises vending taxi trips and hotel stays, among other services. Arguably, very significant gains in distributive service productivity have occurred in the key sense that purchasers’ shopping efficiency has been greatly enhanced by (1) not having to make trips to search physical inventories or negotiate with service providers in disparate locations, and (2) suffering little or no incremental delay (and perhaps an acceleration) in taking delivery of the selected product. By the same token, in 2015 distributive services originated about 12% of US GDP,22 and the advent of the gig economy in particular is not thought to have had much of an impact on GDP growth.23

6.3.5.3.2. Financial Services Measurement of financial services output and intermediate consumption remains an unsettled area, even regarding its nominal value, much less growth in its volume of output and use in intermediate consumption. The national accounts limit financial service output to units in the financial corporations institutional sector, which includes deposit-taking corporations (banks), finance companies, money market investment funds, non-money-market investment funds, and insurance and pension schemes, among others. As investment funds are the most specialized and thus the simplest of financial institutions, we begin with them before discussing output measurement issues with finance companies, banks, and insurance companies. 6.3.5.3.2.1. Investment Funds The nominal output of the portfolio-management services of investment funds is reasonably well understood as a function of the “expense ratio” published by their prospectuses. Shareholders get the return on the fund portfolio minus the product of the expense ratio with the value of (generally financial) assets under management. The expense ratio per unit of currency of assets under management can be rewritten as the difference between the rate of return on the investment portfolio and the rate of return paid out to the investment fund’s shareholders. The latter, return on equity, is self-evidently the investment fund’s cost of capital. So the value of services provided by an investment fund to its funders (who are generally, though not always, all equity holders) per currency unit invested is the rate of return on the fund less the cost of capital or, equivalently, the expense ratio. 6.3.5.3.2.2. Banking and Finance Companies Banking was the initial focus of service imputations at the postwar inception of national accounting standards.24 Some of the output of banks is measured from explicitly charged fees, but the national accounts “indirectly measure” a large part of the output of banks and finance companies as a function of the spread between the income from financial assets and the expense of non-equity liabilities. The national accounts’ “financial intermediation services indirectly measured”

228 Kim Zieschang (FISIM) output and intermediate input measure depends on a “reference rate of interest” and involves the following basic calculation:

(rate of return on financial assets − reference rate of interest) × financial assets + (reference rate of interest − rate of return on non-equity liabilities) × non-equity liabilities. If the reference rate of interest is actually the bank’s or finance company’s cost of capital, then with reference to the investment fund case, FISIM’s first, asset term is self-evidently an asset-management charge in the same vein as the “expense ratio” that investment funds charge their shareholders. If the reference rate is the cost of capital, the second, debt-liability term, applying to banks only (because of their use of deposit financing), is a liquidity service charge. It is the supply value of the liquidity service concept underlying Barnett’s (1980) monetary service aggregates. As of the 2008 version of the national accounting standards, the financial instrument scope of the FISIM output calculation is limited to loans and deposits, which de facto means that FISIM is nearly exclusively applied to banks and finance companies.25 From a productivity point of view, the FISIM liquidity service charge can be broken down into deposit account servicing and liquidity service components, while asset- management FISIM can be decomposed into loan account servicing and portfolio management components. The key volume indicators for the account servicing components on both sides of the ledger is the number of accounts/contracts and number of transactions processed per account/contract.26 Liquidity traditionally has been treated in both the economics and accounting literatures as yielding a volume of services proportional to its “real” or deflated value to the depositor. Asset-management output has been viewed similarly as proportional to the deflated value of assets under management to funders of all types, though portfolio complexity and intensity of research effort, however measured, also appear relevant. Choice of the deflators for debt liabilities and financial assets is seen related to the uses for which a depositor or borrower would spend the funds. Recent economics literature has argued that the FISIM calculation effectively overstates the share of value added in GDP of institutions to which FISIM is applied,27 but this critique remains under discussion. The issue revolves ultimately around what should and should not be considered inside the national accounts “production boundary.” The current FISIM calculation for assets is essentially identical to the “expense ratio” service charge levied by investment funds on their shareholders and that national accountants recognize as an explicit payment for service. The current FISIM calculation for liabilities is a de facto link between financial production in the current national accounts and the liquidity aggregates promulgated in the Divisia aggregation literature in monetary economics.28 At the same time, liquidity services are the key component of FISIM that the recent critics of the established methodology would (de facto) place outside the production boundary, thus treating flows associated with it as a type of appropriation or transfer,

Productivity Measurement in Sectors 229 rather than a payment for service. This contrasts with the monetary aggregation literature (initiated by Barnett 1978, 1980), which considers the demand value of liquidity as payment for a capital service. Intersecting with this connection of the established FISIM principle of national accounting to monetary aggregation theory and practice is the need to distinguish between the production orientation of the national accounts and the use orientation of the monetary aggregation literature. The national accounts attempt to measure liquidity services at their supply value from the enterprise issuers of the debt instruments (deposits, but also debt securities, loans, insurance reserves, and other payables) with which these liquidity services are associated. The monetary aggregation literature measures liquidity at its demand value from the point of view of debt instrument holders. With the supply orientation of the national accounts and the demand orientation of monetary aggregation in mind, there is a reassuring coherence between the evolved measurement practice for the financial service output of banks and finance companies and the measurement of liquidity in the monetary aggregation literature. 6.3.5.3.2.3. Insurance The official measure of current price insurance output is premiums minus claims (net premiums) on policies plus “premium supplements,” the investment income that insurers earn on reserves held for policyholder claims. In 2008 national accountants revised the standards to smooth the often volatile claims component of this nominal output measure. Hornstein and Prescott (1991) proposed a volume indicator for net premiums related to number of policies in force. Their approach tracked output as an index of the premiums less claims of a set of specified policies through time, adjusted as needed using hedonic methods for changes in policyholder and other actuarial characteristics. Reece (1992) provided a reasonably detailed analysis of the price of policy services for life insurance. His approach de facto took the policy contract as the underlying output volume counter. Beginning from first principles, insurers sell policies in risk classes based on policyholder characteristics and other factors relevant to the probability of a claim. Within a risk class, which sets the actuarial estimate of the probability of a claim, insurers set premiums as a percentage per currency unit of coverage. The price of this “vanilla insurance” per currency unit is the premium percentage minus the actuarial claim probability. Because policies are priced as percentages of the face amount insured, however, there remains a fundamentally nominal aspect to vanilla insurance—the face value of the policy. In looking at volume metrics for “vanilla insurance,” clearly there are account servicing aspects to insurance services and, within a broad class of policy, number of policies and number of claims processed should be relevant indicators for the volume of services underlying the net premiums part of national accounts insurance output. By analog with deposits and loans, however, the insured have—and insurers generally require them to have—an “insurable interest” underlying the face value of a policy, in the form of loss or impairment of a nonfinancial durable asset owned by the insured.

230 Kim Zieschang This can be as a result of loss of or damage to a piece of durable equipment or a structure, of a durable consumer asset such as a house or automobile, or of a human capital asset as a result of disability or death. Consequently, a further indicator of insurance service seems to be related to the face value of the policy, deflated by a price index for the insurable interest. It is worth reconsidering what the national accounts term “premium supplements.” Aside from vanilla insurance, there are two other aspects of insurance output not considered explicitly by the national accounts: liquidity and asset-management services. Like banks, insurers finance their operations not only with equity, but also debt. For banks, deposits are the key component of debt financing, and current practice associates provision of liquidity services with deposit debt. For insurers, insurance reserves are the key component of debt financing.29 At the same time, like investment funds, finance companies, and banks, insurers manage a portfolio of financial assets that generate investment income and, like these other financial enterprises, insurers charge their funders an “expense ratio” for that service. While the national accounts exclude the liq uidity and asset-management services associated with financial instruments other than deposit and loan liabilities, productivity analysts should not be proscribed from taking them into account for other types of financial instruments. This granted, the “premium supplements” component of insurance output is better characterized as the sum of the liquidity on insurance reserves that insurers generate in behalf of policyholders and the asset-management services on the total investment portfolio that insurers provide to their equity and debt funders. Volume measures for these two components of insurance output could be developed analogously to the deflation approach used for banks and finance companies for the liquidity and asset-management services they provide to their funders.

6.3.5.3.3. Health Care Health-care service output measurement is straightforward in nominal terms but problematic in volume terms. Scheiner and Malinovskaya (2016) summarize the substantial health-care measurement literature by identifying two related problems with measuring the volume of health-care services: setting a classification for specific health-care services transacted within aggregate health-care expenditure relevant to disease treatment outcomes, and tracking change in the characteristics (quality) of those services once identified. The line of attack for the first problem is to reorganize expenditure and quantity indicator information from a treatment-service classified basis to a disease-classified—medical care expenditure (MCE)—basis. Having selected MCE-basis service classification, the second problem, quality measurement, has been characterized as developing treatment outcome measures by disease. The most easily accessible disease outcome data tend to be for mortality (life expectancy), while comprehensive and regularly reported data on more granular measures of health status remain to be developed or need strengthening. Outcome measurement remains an active area of research, and the quality of medical care remains an unsettled measurement area.

Productivity Measurement in Sectors 231

6.3.5.3.4. Education Education delivery is measured in number of individuals completing primary, tertiary, and secondary levels of study, but these completion rates disguise a range of outcome levels in terms of what graduates actually know. The results of standardized tests of achievement may be important, if still imperfect, additional indicators of educational outcomes.30 For example, as described by the OECD, [t]he Programme for International Student Assessment (PISA) is a triennial international survey which aims to evaluate education systems worldwide by testing the skills and knowledge of 15-year-old students. In 2015 over half a million students, representing 28 million 15-year-olds in 72 countries and economies, took the internationally agreed two-hour test. Students were assessed in science, mathematics, reading, collaborative problem solving and financial literacy.31

National average PISA scores might be used as a quality variable to standardize public and private educational effectiveness between countries. According to Cambridge International Examinations (2015, n.p.), other standardized international testing regimes include TIMSS (Trends in International Mathematics and Science Study), which is repeated every 4 years and tests learners of 10 and 14 years old. It is managed by the International Association for the Evaluation of Educational Achievement (IEA). PIRLS (Progress in International Reading Literacy Study), repeated every 5 years and focusing on 10 year old learners’ abilities in reading, and on national policies concerning literacy.

There are several other testing regimes besides these three. PISA, TIMSS, and PIRLS have the advantage of international standardization, but no regime can yet claim universal and at least annual application, complicating the use of test scores as education service quality variables.

6.4. Concluding Remarks Hard-to-measure sectors display chronic mismeasurement or outright omission of key outputs, intermediate consumptions, and/or capital stocks. Mismeasurement occurs if the period-to-period comparisons of given goods, services, and/or stocks are not of like items; that is, if one or more characteristics of the measured items has changed, or if the number of varieties—the scope—of the types of measured items has changed. This discussion of sectors with “hard to measure” outputs, intermediate consumption, and capital stocks began with a fairly detailed analytical section based on the capacity utilization function, whose dual is the value-added function that underlies

232 Kim Zieschang GDP. We worked with the Translog aggregator and associated Törnqvist index number representations of the capacity utilization function. There are other indexing frameworks, but the Translog/Törnqvist allows exact depiction (to a second-order differential approximation) of multifactor productivity change for a general technology that decomposes into volume factors for output, intermediate consumption, and scale- adjusted capital stock quantity change. Presuming existence of a well-defined (hedonic) functional relationship between the characteristics/quality of outputs, intermediate inputs, and capital stocks in each time period, these volume factors further decompose into quantity factors and associated quality adjustment factors that allow for a changing hedonic locus through time. We characterized the multifactor productivity framework as “semi-nonparametric” because it relies on a parametric form for the hedonic relationship between the prices and associated characteristics of outputs, intermediate inputs, and capital stocks, but does not require parametric estimation of the Translog capacity utilization function.32 Eliminating the need for estimation of technology parameters is an argument for the practicality of this framework, as a good compromise between accurate measurement of multifactor productivity, flexibility of the underlying functional form for technology, and minimal (if not zero) need for estimating structural parameters. As well, its straightforward multiplicative decompositions make it a good heuristic for the practices that economic statisticians actually use in compiling productivity statistics from accounting and other economic data. This semi-nonparametric index number framework is not the only analytical apparatus that might be used for understanding how to chip away at the mismeasurement problem in multifactor productivity. However, it has key flexibilities while remaining reasonably simple. We have discussed some of the restrictions of this framework. Most important, it requires efficiency at all observed levels of prices, quantities, and characteristics, a strong requirement if there is reason to suppose there is significant scale inefficiency in our observations. Allowing for inefficiency seems to require that we know more about the parametric structure of technology than superlative index numbers allow, as suggested by O’Donnell (2015). On the other hand, there is a great deal not known about the evolution of multifactor productivity that is related to unmeasured or imperfectly measured quantities of some outputs/intermediate consumptions, capital stocks, and, important, the changing characteristics or quality of these key productivity drivers. The latter may well be at least as important as disallowing scale inefficiency, and this framework fixes a spotlight on the importance of dealing with mismeasurements and omitted variables in assessing the evolution of an inherently residual concept like multifactor productivity. We briefly considered issues in the main hard-to-measure productive activities. The cited references in that discussion contain lists of additional sources that go more deeply into the sector-specific issues. Among key frontier measurement challenges are those involving the contributions of human and environmental capital to production. Although labor services metrics are available in most countries and are accounted for in standard productivity measures, human capital assets, while likely to comprise the principal part of national wealth in many countries, are not included within the national

Productivity Measurement in Sectors 233 accounts asset boundary. However, much is known about human capital, and statistical standards for including it in official output and wealth are in process.33 Although the asset boundary of existing national accounting standards includes natural resources such as subsoil mineral assets and cultivated biological assets, the depletion of these assets is not accounted for in existing measures of net national output, where it plays a similar role to depreciation of produced capital assets. International standards have been developed to measure depletion, but official measures remain under development.34 Beyond natural resources in measuring the contribution of natural capital to output, however, are ecosystems and the associated ecosystem services, which currently lie outside the asset boundary of the national accounts due in part to significant measurement challenges.35 There is much being done, but there is a great deal more to be done in deciding what concepts to measure and, once decided, secularly measuring those concepts.

Notes 1. The endogenous technical progress literature aims to more fully close the specification of technology and more fully explain productivity and efficiency residuals as the resultants of the accumulation of excluded or poorly measured capital inputs, namely, nonrival and partially or fully nonexcludable knowledge assets. In recognition of the role of knowl edge capital, national accounting standards have been revised (European Commission, IMF, OECD, UN, and World Bank (2008), paragraph 1.38, p. 555) to include a new class of nonfinancial asset—Intellectual property products (AN117)—including

•

Research and development (AN1171) Mineral exploration and evaluation (AN1172) • Computer software and databases (AN1173) • Entertainment, literary or artistic originals (AN1174) • Other intellectual property products (AN1179). •

Progress since 2008 has been steady in developing data sources and statistical measures for these assets, but much remains to be done. 2. In the interest of simplicity, this chapter assumes that the various maxima and minima defining the various distance function representations of a joint production function are attainable on the technology set T , and subsume “defined in the limit” (“infimum” or “supremum”) into “minimum” or “maximum.” If these minima and maxima are not bounded, we say the resultant is undefined. 3. Shephard’s (1953, 1970) definition of the input distance function is the reciprocal of the input distance function defined here. 4. Balk (2009) employs the capacity utilization concept. 5. However, because we do not assume the technology T is convex, the converse— if d yx ( y , x , K , γ , χ, κ , ϑ ) ≤ 1 then ( y , x , K , γ , χ.κ , ϑ ) is feasible—is not necessarily true. 6. We are treating labor as a human capital stock, rather than a rented service, to highlight its role as a quasi-fixed factor of production like nonhuman capital. The standard national accounts presentation does not account for human capital, only the rental of it in the form of “compensation of employees.” On the other hand, the presence of nonhuman capital

234 Kim Zieschang

in stock form among primary inputs does imply here that the using enterprise owns the nonhuman asset. If the nonhuman capital asset is not owned by the enterprise, then as in the most recent, 2008 version of national accounting standards, it enters the production function of that enterprise as an intermediate consumption of rented services and thus is not in value added. 7. Observed states are indicated in equation (6.1) when the superscript on the capacity utilization function and the superscripts on all of its arguments match, as in d 0yx ( y 0 , x 0 , K 0 , γ 0 , χ0 , κ 0 , ϑ 0 ) and d1yx ( y1 , x1 , K 1 , γ 1 , χ1 , κ1 , ϑ1 ). Non- observed states are indicated when the superscript on the capacity utilization function and/or the superscripts on its arguments differ, as in d 0yx ( y1 , x1 , K 1 , γ 1 , χ1 , κ1 , ϑ 0 ) and d1yx ( y 0 , x 0 , K 0 , γ 0 , χ0 , κ 0 , ϑ1 ). When observed states are efficient, d 0yx ( y 0 , x 0 , K 0 , γ 0 , χ0 , κ 0 , ϑ 0 ) = 1 and d1yx ( y1 , x1 , K 1 , γ 1 , χ1 , κ1 , ϑ1 ) = 1. 8. Basic prices in national accounts are the prices the producer receives for output; they exclude taxes on products, include subsidies on products, and exclude separately invoiced transportation and distribution margins. 9. Purchasers’ prices are the prices the producer pays for input; they include taxes on products, exclude subsidies on products, and include separately invoiced trade and transportation margins. 10. If a = (a1 , a2 ,, an )′ and b = (b1 , b2 ,, bn )′ their Hadamard product is a b = (a1b1 , a2b2 , …, anbn ). 11. Being subject to eventual decreasing returns to scale means that it is always possible to find outputs, intermediate consumptions, and capital stocks sufficiently large that ε ds yx ′K =

K s′ ∇ K d y ,x dθ ≤ 1. It does not preclude increasing returns to scale ε ds yx′K > 1 = − s′ dλ y ∇ y d y ,x + x s′ ∇ x d y ,x

at lower levels of outputs and intermediate consumptions, as exemplified by S-shaped plots of the scale of outputs and intermediate consumptions against the scale of primary inputs. 12. We use the term “super surplus” to distinguish the margin p s′ y s − v s′ x s − u s′ K s from the national accounts’ “operating surplus” p s′ y s − v s′ x s − uLs′ K Ls. Neither of these is “profit” as understood in commercial or national accounting (in national accounts “profit” goes by the term “entrepreneurial income”), which includes, among other things, the income and expense from the financial instruments the enterprise owns or for which it is liable that is not attributable to a financial service flow. 13. Dealing with the elasticity of scale becomes rather slippery in application. According to s ′ s CCD(1982), if s us K s s > 1, (corresponding to negative super surplus) and/or returns p ′ y − v ′x

to scale are known to be increasing at one or both observed states, then we have to estimate ε ds

yx ,K

by other means. On the other hand, when

us′K s ≤ 1, (corresponding to ps′ y s − v s′ x s

nonnegative super surplus), CCD(1982) note that we can determine the elasticity of scale as ε ds

yx ′ K

=

us′ K s ≤ 1; that is, returns to scale are nonincreasing. Intermediate microecps′ y s − v s′ x s

onomic theory associates the long run equilibrium of an enterprise varying all inputs, including quasi fixed inputs, with operating at the point of unitary returns to scale, ε ds ,K = 1. This latter case corresponds to zero super surplus (thus, we could argue that super surplus is, de facto, the “economic profit” of microeconomics). Under the assumption that yx

Productivity Measurement in Sectors 235 the enterprise is operating at a point of unitary returns to the scale of quasi fixed (primary) inputs K s , or long run equilibrium, ε ds ,K disappears from the multifactor productivity index (6.2). Allowing for the possibility of nonunitary returns to scale in applications of this framework to accounting data would involve specific treatments of the user cost of quasi fixed capital stocks that do not imply that super surplus is zero. 14. Feenstra (1995) derived a similar, though not identical, result for quality- adjusted Törnqvist household consumption price indices. 15. This is particularly true of land assets, whose prices are most directly affected by environmental quality characteristics. 16. Jorgenson (1966) attributed the “disembodied” technical progress concept to Tinbergen (1959) and its complement—“embodied” technical progress—to Solow (1960). Hulten (1992) observed that technical change embodied in inputs (e.g., capital stocks) can be equivalently viewed as an adjustment for the quantity of inputs or to the price of those inputs. 17. On intellectual property assets, see European Commission, IMF, OECD, UN, and World Bank (2008), paragraphs 10.98–10.117, 206–207. 18. On goodwill and marketing assets, see European Commission, IMF, OECD, UN, and World Bank (2008), paragraphs 10.196–10.199, 216. Prevailing commercial and naöional accounting standards allow booking of goodwill and marketing assets only if their value has been revealed in a previous transaction, thus the “purchased goodwill and marketing assets” terminology in the aforementioned 2008 SNA reference. 19. The usefulness of productivity as a measure of ignorance, or lack thereof, in specific contexts is arguably relevant. For example, Paul Romer (2016) recently has critiqued “real business cycle” macroeconomic models by asking why fluctuations in a measure of our ignorance (“productivity shocks,” which he terms “phlogiston”) should be a key independent variable in a macroeconomic model. 20. On reservation prices for new and disappearing products, see Fisher and Shell (1972). Some indicative exact results for new and disappearing goods in the vein of the translog “semi- parametric” results in this chapter are in a rather obscure working paper by Zieschang (1989). Lovell and Zieschang (1994) approached the reservation price problem by estimating shadow prices via linear envelopment (data envelopment analysis or DEA). See Feenstra (1995) for exact index number results for new and disappearing products when the aggregator (e.g., cost or revenue) function has constant elasticity of substitution (CES) form. 21. Eurostat, International Labour Organisation, International Monetary Fund, Organisation for Economic Cooperation and Development, United National Economic Commission for Europe, and World Bank (2013). 22. Bureau of Economic Analysis, Annual Industry Accounts (GDP by Industry and Input- Output Accounts), https://bea.gov; Strassner and Wasshausen (2014). 23. As noted at the beginning of this section, Bryne et al. (2016) did not find large impacts on US GDP growth from improving measurement of the high technology sector, including web-enabled distributive services. Nakamura, Samuels, and Soloveichik (2016) reach a similar conclusion. This said, measurement in web-enabled distribution remains an unsettled area. 24. Stone (1947, 40–41) and United Nations (1953, 39). 25. European Commission, International Monetary Fund, Organisation for Economic Co- operation and Development, United Nations, and World Bank (2008). The previous, yx

236 Kim Zieschang 1993 version of the same standards did not proscribe FISIM to deposits and loans, thus implicitly allowing the scope of FISIM to be all financial instruments, though it did exclude “own funds” (equity) from playing a role in FISIM production. The 1993 national accounting standards thus did not foreclose alignment of the liquidity services component of FISIM with the liquidity services recognized in the monetary aggregation literature, which includes liquidity generated not only from deposits but from almost all other forms of debt, in particular debt securities (bonds). 26. The US Bureau of Labor Statistics takes this approach to account servicing-type products in US price and productivity statistics, for example. See Royster (2012). 27. Basu, Inklaar, and Wang (2011); Colangelo and Inklaar (2012); Inklaar and Wang (2013). 28. Barnett (1978, 1980), Donovan (1978), and the large economics literature that they started. 29. See, e.g., Task Force on Financial Statistics (2013, 29, paragraph 3.3) for the financial instruments classified as debt, among which are insurance reserves. 30. Hanushek and Ettema (2016). 31. http://www.oecd.org/pisa/aboutpisa/. 32. Or the dual of the capacity utilization function, the value-added function. The value- added function is a form of restricted profit function. 33. United Nations Economic Commission for Europe Task Force on Measuring Human Capital (2016). 34. Eurostat, International Monetary Fund, Organisation for Economic Co-operation and Development, United Nations, United Nations Food and Agriculture Organisation, World Bank (2012). 35. European Commission, Organisation for Economic Co-operation and Development, United Nations, World Bank (2013).

References Balk, Bert M. 2009. “On the Relation Between Gross Output-and Value Added-Based Productivity Measures: The Importance of the Domar Factor.” Macroeconomic Dynamics 13(S2) (Measurement with Theory): 241–267. Barnett, William A. 1978. “The User Cost of Money.” Economics Letters 1(2): 145–149. Barnett, William A. 1980. “Economic Monetary Aggregates: An Application of Index Number and Aggregation Theory.” Journal of Econometrics 14(1): 11–48. Basu, Susanto, Robert Inklaar, and Christina Wang. 2011. “The Value of Risk: Measuring the Service Output of U.S. Commercial Banks.” Economic Inquiry 49(1): 226–245. Bosworth, Barry, and Jack Triplett. 2007. “Services Productivity in the United States; Griliches’s Services Volume Reconsidered.” In Hard-to-Measure Goods and Services: Essays in Honor of Zvi Griliches, edited by Barry Bosworth and Jack Triplett, 413–447. Chicago: Chicago University Press for National Bureau of Economic Research. Bureau of Economic Analysis. Various quarters. Annual Industry Accounts (GDP by Industry and Input-Output Accounts). https://bea.gov. Byrne, David M., John G. Fernald, and Marshall B. Reinsdorf. 2016. “Does the United States Have a Productivity Slowdown or a Measurement Problem?” Brookings Papers on Economic Activity (Spring): 109–157. Cambridge International Examinations. 2015. “International Surveys: PISA, TIMSS, PIRLS.” Education Brief 7. http://www.cie.org.uk/images/271193-international-surveys-pisa-timss- pirls.pdf.

Productivity Measurement in Sectors 237 Cole, Rosanne, Y. C. Chen, Joan A. Barquin-Stollerman, Ellen Dulburger, Nurhan Halvacian, and James H. Hodge. 1986. “Quality-Adjusted Price Indexes for Computer Processors and Selected Peripheral Equipment.” Survey of Current Business 66(1): 41–50. Colangelo, Antonio, and Robert Inklaar. 2012. “Bank Output Measurement in the Euro Area: A Modified Approach.” Review of Income and Wealth, Series 58, 1 (March): 142–165. Donovan, Donal J. 1978. “Modeling the Demand for Liquid Assets: An Application to Canada.” IMF Staff Papers 25 (December), 676–704. Dulberger, Ellen R. 1989. “The Application of a Hedonic Model to a Quality-Adjusted Price Index for Computer Processors.” In Technology and Capital Formation, edited by Dale W. Jorgenson and Ralph Landau. Cambridge, MA: MIT Press. Caves, Douglas W., Laurits R. Christensen, and W. Erwin Diewert. 1982. “The Economic Theory of Index Numbers and the Measurement of Input, Output, and Productivity.” Econometrica 50(6): 1393–1414. European Commission, International Monetary Fund, Organsation for Economic Co- operation and Development, United Nations, United Nations Food and Agriculture Organisation, and World Bank. 2012. System of Environmental-Economic Accounting 2012 Central Framework. https://unstats.un.org/unsd/envaccounting/seeaRev/SEEA_CF_Final_ en.pdf. European Commission, International Monetary Fund, Organisation for Economic Co- operation and Development, United Nations, and World Bank. 1993. System of National Accounts 1993. New York: United Nations. European Commission, International Monetary Fund, Organisation for Economic Co- operation and Development, United Nations, and World Bank. 2008. System of National Accounts 2008. New York: United Nations. European Commission, Organisation for Economic Co-operation and Development, United Nations, and World Bank. 2013. System of Environmental- Economic Accounting 2012 Experimental Ecosystem Accounting. https://unstats.un.org/unsd/envaccounting/eea_ white_cover.pdf. Eurostat, International Labour Organisation, International Monetary Fund, Organisation for Economic Co-operation and Development, United National Economic Commission for Europe, and World Bank. 2013. Handbook on Residential Property Prices Indices (RPPIs), http://ec.europa.eu/eurostat/documents/3859598/5925925/KS-RA-12-022-EN.PDF. Feenstra, Robert C. 1995. “Exact Hedonic Price Indexes.” The Review of Economics and Statistics 77(4): 634–653. Fisher, F. M., and K. Shell. 1972. The Economic Theory of Price Indices. New York: Academic Press. Fisher, Irving. 1922. The Making of Index Numbers: A Study of Their Varieties, Tests, and Reliability. New Haven, CT: Houghton Mifflin. Fixler, Dennis, and Kimberly Zieschang. 1992. “Incorporating Ancillary Measures of Process and Quality Change into a Superlative Productivity Index.” Journal of Productivity Analysis 2: 243–267. Graf, Brian, and Mick Silver. 2014. “Commercial Property Price Indexes: Problems of Sparse Data, Spatial Spillovers, and Weighting.” IMF Working Paper 14/72. https://www.imf.org/~/ media/Websites/IMF/imported-full-text-pdf/external/pubs/ft/wp/2014/_wp1472.ashx. Hanushek, Eric A., and Elizabeth Ettema. 2016. “Defining Productivity in Education: Issues and Illustrations.” The American Economist 65(2), October 2017, 165–183. Hornstein, Andreas, and Robert C. Prescott. 1991. “Measures of the Insurance Sector Output.” The Geneva Papers on Risk and Insurance 16 (59): 191–206.

238 Kim Zieschang Hulten, Charles R. 1992. “Growth Accounting When Technical Change Is Embodied in Capital.” The American Economic Review 82(4): 964–980. Inklaar, Robert, and Christina Wang. 2013. “Real Output of Bank Services: What Counts Is What Banks Do, Not What They Own.” Economica, London School of Economics and Political Science, 80(317): 96–117. Jorgenson, Dale W. 1963. “Capital Theory and Investment Behavior.” The American Economic Review 53(2), Papers and Proceedings of the Seventy-Fifth Annual Meeting of the American Economic Association (May): 247–259. Jorgenson, Dale W. 1966. “The Disembodiment Hypothesis.” Journal of Political Economy 1(Feb): 1–17. Laspeyres, E. 1871. :Die Berechnung einer mittleren Waarenpreissteigerung.” Jahrbücher für Nationalökonomie und Statistik 16: 296–314. Lovell, C. A. Knox, and Kimberly Zieschang. 1994. “The Problem of New and Disappearing Commodities in the Construction of Price Indexes.” In Data Envelopment Analysis: Theory, Methodology, and Applications, edited by Abraham Charnes, William W. Cooper, Arie Y. Lewin, and Lawrence M. Seiford, 353–367. Dordrecht: Springer. Nakamura, L., J. Samuels, and R. Soloveichik. 2016. “Valuing ‘Free’ Media in GDP, 1929–2015.” Paper presented at the Allied Social Science Associations meeting, January 8. O’Donnell, C. J. 2015. “Using Information about Technologies, Markets and Firm Behaviour to Decompose a Proper Productivity Index.” Journal of Econometrics 190: 328–340. Paasche, H. 1874. “Über die Preisentwicklung der letzten Jahre nach den Hamburger Borsennotirungen.” Jahrbücher für Nationalökonomie und Statistik 12: 168–178. Reece, William S. 1992. “Output Price Indexes for the U.S. Life Insurance Industry.” The Journal of Risk and Insurance 59(1): 104–115. Romer, Paul M. 1990. “Endogenous Technological Change,” Journal of Political Economy 98(5), Part 2: The Problem of Development: A Conference on the Institute for the Study of Free Enterprise Systems (October): S71–102. Romer, Paul M. 2016. “The Trouble With Macroeconomics,” https://paulromer.net/wp-content/ uploads/2016/09/WP-Trouble.pdf. Royster, Sara E. 2012. “Improved Measures of Commercial Banking Output and Productivity.” Monthly Labor Review (July): 3–17. Scheiner, Louise, and Anna Malinovskaya. 2016. “Measuring Productivity in Healthcare: An Analysis of the Literature, Hutchins Center on Fiscal and Monetary Policy.” Washington, DC: Brookings Institution. https://www.brookings.edu/wp-content/uploads/2016/08/hp- lit-review_final.pdf. Shephard, Ronald W. 1953. Cost and Production Functions. Princeton, NJ: Princeton University Press. Shephard, Ronald W. 1970. Theory of Cost and Production Functions. Princeton, NJ: Princeton University Press. Solow, Robert M. 1960. “Investment and Technical Progress.” In Mathematical Methods in the Social Sciences, edited by K. J. Arrow, S. Karlin, and P. Suppes, 89–104. Stanford, CA: Stanford University Press. Stone, Richard. 1947. “Appendix: Definition and Measurement of the National Income and Related Totals.” In Measurement of National Income and the Construction of Social Accounts: Report of the Sub-Committee on National Income Statistics of the League of Nations Committee of Statistical Experts. http://unstats.un.org/unsd/nationalaccount/docs/ 1947NAreport.pdf.

Productivity Measurement in Sectors 239 Strassner, Eric, and David B. Wasshausen. 2014. “New Quarterly Gross Domestic Product by Industry Statistics.” BEA Briefing, Bureau of Economic Analysis. https://www.bea.gov/scb/ pdf/2014/05%20May/0514_gdp-by-industry.pdf. Sveikauskas, Leo, Samuel Rowe, James Mildenberger, Jennifer Price, and Arthur Young. 2014. “Productivity Growth in Construction.” BLS Working Paper 478. https://www.bls.gov/ osmr/pdf/ec140090.pdf. Task Force on Financial Statistics. 2013. External Debt Statistics: Guide for Compilers and Users. Washington, DC: International Monetary Fund. http://www.tffs.org/edsguide.htm. Tinbergen, J. 1959. “On the Theory of Trend Movements.” In Jan Tinbergen Selected Papers, edited by L. H. Klaassen, L. M. Koyck, and H. J. Witteveen, 182–221. Amsterdam: North-Holland (originally published in German as “Zur Theorie der langfristigen Wirtschaftsentwicklung,” Weltwirtschaftliches Archiv LV(1) [1942]: 511–549). United Nations Economic Commission for Europe Task Force on Measuring Human Capital. 2016. Guide on Measuring Human Capital. https://unstats.un.org/unsd/nationalaccount/ consultationDocs/HumanCapitalGuide%20Global%20Consultation-v1.pdf. Zieschang, Kimberly. 1989. “The Characteristics Approach to the Problem of New and Disappearing Goods in Price Indexes.” BLS Working Paper WP-183. US Bureau of Labor Statistics. https://www.bls.gov/osmr/workpapers_catalog1992.htm#1992.

Chapter 7

Produ ct i v i t y Measu rem e nt i n t he Public Se c tor W. Erwin Diewert

7.1. Introduction: How Should Public-Sector Non-M arket Outputs Be Valued? In order to measure the total factor productivity (TFP) of a public-sector production unit (or establishment) using index number techniques, it is necessary to measure the prices and quantities of the outputs produced and the inputs used by that unit for two periods of time.1 Then TFP growth can be defined as a quantity index of outputs produced, divided by a quantity index of inputs used by the establishment.2 It is usually possible to measure the price and quantity of inputs in a fairly satisfactory manner,3 but there are problems in measuring the prices and quantities of public-sector nonmarket outputs. Thus in this chapter, we will take a systematic look at possible methods for the valuation of non-market outputs produced by public-sector production units. In many cases, it is difficult to determine exactly what it is that a public-sector production unit produces. In this case, we may have neither quantities nor prices for the outputs of the government service provider. However, in many cases, we can measure at least the quantities of the outputs produced by the public-sector unit but not the corresponding prices. Finally, in some cases, it may be possible to measure output quantities produced and to obtain estimates of purchaser’s valuations for the missing non-market output prices. Thus the “best practice” methodology that can be used to form productivity estimates for the public-sector unit will depend to a large extent on what information on prices and quantities is available.

242 W. Erwin Diewert From the perspective of measuring the effects on the welfare of households of public- sector production, we suggest the following methods for valuing government outputs in the order of their desirability: • First best: valuation at market prices or purchaser’s valuations; • Second best: valuations at producer’s unit costs of production; • Third best: output growth of the public-sector production unit is set equal to real input growth, and the corresponding output price growth is set equal to an index of input price growth. Obviously, the third-best option is the least desirable option. If it is used, then productivity growth for the public-sector production unit will be nonexistent by construction. If there are competitive markets with no economies of scope and constant returns to scale in production, then the first-and second-best options will be roughly equivalent (i.e., the purchaser’s price will be approximately equal to the long-run marginal cost of producing a unit of the commodity). However, in a non-market setting, even with no economies of scope and constant returns to scale in production, there is nothing to force the cost of producing a unit of public-sector output to equal its value to recipients of the commodity.4 In an extreme case, the production unit could be producing nothing of value. Thus from the viewpoint of welfare economics, valuation of public-sector outputs at purchasers’ valuations appears to be the preferred option.5 The task of this chapter is to suggest methods for measuring the TFP of a public- sector production unit that produces at least some outputs that are allocated to recipients at non-market prices (i.e., at zero prices or at highly subsidized prices that do not cover their unit costs of production). In general, the TFP level of a production unit is defined as the real output produced by the production unit at a time period divided by the real input utilized by the production unit during the same time period. There is TFP growth if real output grows more rapidly than real input. The main drivers of TFP growth are the following: (i) technical progress (i.e., at outward shift of the unit’s production possibilities set); (ii) increasing returns to scale combined with input growth; and (iii) improvements in technical and allocative efficiency.6 This volume outlines many methods for measuring TFP, but in this chapter, we will concentrate on index number methods.7 The use of index number methods means that we will not be able to decompose the TFP growth of a public-sector production unit into the contributions of the preceding three main components of TFP growth (i.e., index number methods do not allow us to measure separately the effects of these explanatory factors; in general, all three explanatory factors will be combined into the overall TFP growth measure). However, the use of index number methods requires a methodological framework for the determination of output prices for the non-market outputs produced by the public sector. If the public-sector unit minimizes its cost of production in producing its non- market outputs, then from the viewpoint of the economic approach to productivity measurement, the “right” prices to use to value non-market outputs are the (long-run) marginal costs of producing the non-market outputs. Diewert (2012, 222–228) provided

Productivity Measurement in the Public Sector 243 a nonparametric methodological justification for this marginal cost method of output price valuation for non-market outputs.8 Diewert also justified the use of the Fisher (1922) index number formula to aggregate inputs and outputs using his approach. We will summarize his approach in the Appendix to this chapter. Variants of this cost- based method for valuing non-market outputs will be discussed in section 7.5. The third-best option outlined earlier is the only option that can be used when there is little or no information on both the prices and quantities produced by a government establishment. This is the option that is recommended in the System of National Accounts 1993 to value government production when direct information on the prices and quantities of government outputs is not available. The quantity or volume measure for establishment output that results from using this methodology can be interpreted as a measure of real resources used by that establishment, and as such, it is an acceptable indicator of the output produced by a government unit. This third-best option will be discussed in more detail in the following three sections. Section 7.2 introduces the method. Section 7.3 discusses how user costs should be used to value public-sector capital stock inputs. Section 7.4 addresses the problems associated with choosing an index number formula and how TFP could be measured either as gross output TFP or as value-added TFP. This section also shows that the choice of an index number formula to aggregate outputs and inputs does matter. The second-best option for valuing public-sector outputs will be discussed in section 7.5 and the first-best option in section 7.6. Section 7.7 provides a numerical example due to Schreyer (2010, 21) that illustrates how the first-best option (from the viewpoint of welfare economics) could be used in order to value non-market outputs. This section also discusses some aspects of the problem of adjusting output quantities for quality change. Section 7.8 discusses some of the practical difficulties associated with the measurement of public-sector outputs in selected industries. Section 7.9 concludes.

7.2. The Case Where No Information on the Prices and Quantities of Non-M arket Outputs Is Available The third-best option outlined earlier is the only option that can be used when there is little or no information on both the prices and quantities produced by a public-sector production unit (or there is no agreement on how to measure the outputs of the unit). For example, usually there is little information on the price and quantity of educational services produced by the public sector.9 As was mentioned earlier, the quantity or volume measure for establishment output that results from using this methodology can be interpreted as a measure of real resources used by the production unit, and as such, it is an acceptable indicator of the output produced by the unit. Although this third-best option is fairly straightforward in principle (and has been extensively discussed in the

244 W. Erwin Diewert national income accounting literature), there are some aspects of the method that deserve some additional discussion. The two aspects of the third method that we will discuss in more detail in the following two sections are as follows: • How exactly should the contribution of durable inputs used in a public-sector production unit to current period production be valued? • How exactly should estimates for the aggregate real output produced by a non- market production unit be constructed? In particular, which index number formula should be chosen to perform the aggregation, and does the choice of formula make a difference?

7.3. The Valuation of Durable Inputs Used in the Public Sector The basic tool that economists use to value the contribution of a durable input (or a capital input) to production in an accounting period is the concept of a user cost. We will first explain how to construct the user cost of a capital input10 for a durable input for which market prices exist for the same input at different ages. Consider a production unit that purchases qt units of a durable input at the beginning of accounting period t at the price Pt. After using the services of the capital input during period t, the production unit will have qu t = (1 − δ)qt units of used or depreciated capital (in constant quality units) on hand at the end of period t where δ is the one period depreciation rate for the capital good under consideration.11 Finally, we assume that the production unit has a one period financial opportunity cost of capital at the beginning of period t (i.e., a beginning of the period nominal interest rate) equal to rt. The gross cost of the capital input is the beginning of the period purchase cost of the capital inputs, P t qt . But this cost is offset by the revenue that could be raised by selling the depreciated capital stock value at its imputed market value at the end of the period, which is P t +1qu t = P t +1 (1 − d )qt . But this imputed revenue is not equivalent to the cost outlay made at the beginning of the period. To make it equivalent, we need to take into account that money received at the end of the period is less valuable than money received at the beginning of the period, and so the end of period market value should be discounted by 1+ r t . Thus the net cost of using the services of the capital input during period t is Ut, defined as follows:12

(

(

U t ≡ Pt − 1 + rt

)

−1

(1 − d )P t +1 .

)

(7.1)

Define the constant quality asset inflation rate over period t, it, by the following equation:

1 + i t ≡ PK t +1 / PK t . (7.2)

Productivity Measurement in the Public Sector 245 Substituting (7.2) into (7.1) leads to the following expression for the ex post user cost Ut defined by (7.1):

(

) (1 − δ ) (1 + i ) P

(

)

(

)

= (1 + r t )−1 1 + r t − (1 − d ) 1 + i t P t . (7.3) Rather than discounting the end-of-period value of the capital stock to the beginning of the period, it is more convenient to anti-discount costs and benefits to the end of the period. Thus the ex post end of period user cost of a capital input, ut, is defined as 1+ r t times Ut:13 U t ≡ Pt − 1 + rt

−1

t

t

(

(

)

(

)

(

)

(

)

)

ut ≡ 1 + r t U t = 1 + r t − (1 − δ ) 1 + i t P t = r t − i t + δ 1 + i t P t . (7.4)

Using this formula, we add the opportunity cost of tying up financial capital for one period, rtPt, to the beginning of the period purchase cost of the asset, Pt, to obtain a total cost of purchasing one unit of an asset and tying up financial capital for one period, which is 1+ r t P t . This total asset cost is offset by the end-of-period (imputed) benefit of the depreciated value of one unit of the purchased capital stock, (1 − d ) 1 + i t P t . Taking the difference of this cost and benefit leads to the user cost formula (7.4). This ex post formula for the user cost of capital defined was obtained by Christensen and Jorgenson (1969, 302) for the geometric model of depreciation.14 However, the derivation of the user cost of capital defined by (7.3) or (7.4) is not the end of the problems associated with valuing the cost of using a capital input over an accounting period. There are four additional issues that need to be discussed:

(

• • • •

)

(

)

Ex ante versus ex post user costs; How to treat specific taxes on some capital inputs such as property taxes; What to do if there are no market prices for used capital stocks; and What should be done if inappropriate user costs of capital are used.

We will address each of these problems in turn. The problem with the user cost formulae defined by (7.3) or (7.4) is that they use the ex post or actual asset inflation rate it defined by (7.2). For land assets in particular, ex post asset inflation rates can at times be so large that it makes the user costs defined by (7.3) or (7.4) negative. Basically, we would like the user cost of an asset to be approximately equal to the rental price for the asset (if rental markets for the asset exist). Rental prices will rarely be negative, so obviously a negative user cost will be a poor approximation to a rental price. A way around the negative user cost problem is to replace the ex post asset inflation rate by an anticipated asset inflation rate. Thus suppose that at the beginning of period t, the anticipated end of period t price for an asset of the same quality is P t +1*. Use this price in order to define an anticipated asset inflation rate as i t * ≡ P t +1* / P t − 1. Now replace the ex post asset inflation rate it that appears in the user cost formulae (7.3) and (7.4) by their anticipated counterparts i t * and we obtain the corresponding ex ante user costs, U t * and ut * . Jorgenson (1989, 1996) and his coworkers15 endorsed the use of ex post user costs, arguing that producers can perfectly anticipate future asset prices.

(

)

246 W. Erwin Diewert On the other hand, Diewert (1980, 476; 2005, 492–493), Schreyer (2001, 2009), and Hill and Hill (2003) endorsed the ex ante version for most purposes, since these ex ante user costs will tend to be smoother than their ex post counterparts, and they will generally be closer to a rental or leasing price for the asset.16 Diewert and Fox (2016) used sectoral data on the US corporate and non-corporate financial sector to compute capital services aggregates and the resulting rates of TFP growth using both Jorgensonian and smoothed user costs that use predicted asset inflation rates.17 They found that Jorgensonian ex ante user costs for land components were indeed negative for many years, but the use of ex ante or predicted asset inflation rates cured the problem of negative user costs and, indeed, led to much smoother user costs, as could be expected.18 There is another solution to the problem of negative user costs if rental prices for the asset are available: take the maximum of the user cost and the corresponding market rental price as the appropriate valuation for the services of the asset during the period under consideration. The user cost valuation for the services of the asset is essentially a financial opportunity cost of using the asset, while the rental price is the opportunity cost of using the services of the asset for productive purposes, rather than renting it out for the period. Both valuations are valid opportunity costs, so the true opportunity cost of using the asset should be the maximum of these two costs. Diewert (2008) called this the opportunity cost method for asset services valuation.19 We have ignored tax complications in deriving the user cost formulae (7.3) and (7.4). Any specific capital taxes (such as property taxes on real estate assets) should be added to the user cost formula for the relevant assets.20 Business income taxes that fall on the gross return to the asset base can be absorbed into the cost of capital, rt, so that rt can be interpreted as the before income tax gross return to the asset used by the production unit.21 The user costs (7.3) and (7.4) were derived under the assumption that there are market prices for used assets that can be used to value the asset at the end of the accounting period. However, for many unique assets that do not trade in each accounting period (such as real estate, intellectual property, and mining assets, and certain types of artistic assets, such as a movie), there are no end-of-period asset prices that are available. In these cases, estimates of the future discounted cash flows that the asset might generate have to be used in order to value the asset as it ages. It will usually be difficult to form these estimates.22 There is still a certain amount of controversy on how exactly to measure capital services in the international System of National Accounts (SNA).23 A particular problem with the System of National Accounts 1993 and System of National Accounts 2008 is that capital services in the general government sector are to be measured by depreciation only; that is, there is no allowance for the opportunity cost of capital in government- sector user costs,24 whereas as we have seen in the preceding, market-sector user costs of capital include both depreciation and the opportunity cost of capital that is tied up in holding productive assets. This omission of imputed interest cost will lead to a substantial underestimate of public-sector costs (from an opportunity cost perspective) and hence economy wide gross domestic product (GDP) will also be underestimated.25

Productivity Measurement in the Public Sector 247 The Office of National Statistics (ONS) in the United Kingdom has made a substantial effort to measure productivity in the public sector, and it recognized that the SNA- recommended treatment of capital input in the public sector is not appropriate for productivity measurement purposes. Thus the ONS treatment of capital services costs in the public sector for productivity measurement purposes is different from its SNA treatment, which follows the international guidelines. This is unfortunate because ideally, we would like the official GDP measure to coincide with the GDP measure that is used for productivity measurement purposes. The preceding material should alert the reader to the fact that the measurement of capital services input in the public sector is not a completely straightforward exercise. Thus it is not that simple to measure the non-market output produced by a public-sector production unit by its corresponding input measure due to the fact that it is not completely straightforward to measure capital services in both the private and public sector.

7.4. Measuring Output Growth by Input Growth: Gross Output versus Value Added If it proves to be difficult or impossible to measure non-market output quantities, then as indicated earlier, economic statisticians have generally measured the value of non- market outputs by the value of inputs used and implicitly or explicitly set the price of non-market output equal to the corresponding input price index. Atkinson (2005, 12) describes the situation in the United Kingdom prior to 1998 as follows: In many countries, and in the United Kingdom from the early 1960’s to 1998, the output of the government sector has been measured by convention as the value equal to the total value of inputs; by extension the volume of output has been measured by the volume of inputs. This convention regarding the volume of government output is referred to below as the (output = input) convention, and is contrasted with direct measures of government output. The inputs taken into account in recent years in the United Kingdom are the compensation of employees, the procurement costs of goods and services and a charge for the consumption of fixed capital. In earlier years and in other countries, including the United States, the inputs were limited to employment.

As was noted in the previous section, the preceding conventions imply that capital services input for government-owned capital will generally be less than the corresponding capital services input if the capital services were rented or leased. In the owned case, the government user cost of capital consists only of depreciation, but in the leased case, the rental rate would cover the cost of depreciation plus the opportunity cost of the financial

248 W. Erwin Diewert capital tied up in the capital input. Atkinson (2005, 49) makes the following recommendation on this issue (and we concur with his recommendation): We recommend that the appropriate measure of capital input for production and productivity analysis is the flow of capital services of an asset type. This involves adding to the capital consumption an interest charge, with an agreed interest rate, on the entire owned capital.

In addition to the preceding problem, there are some subtle problems associated with measuring public-sector outputs by their real input utilization: • How exactly should real input be measured (i.e., which index number formula should be used to aggregate inputs, and does the choice of formula matter)? • Should the output aggregate for a public-sector production unit be measured by its real primary input or by its real gross input (i.e., by primary plus intermediate inputs)? In order to address these questions, we will calculate alternative input aggregates for an artificial data set. In Table 7.1, we list the prices w1t , w2t and the corresponding quantities x1t , x2t for two primary inputs and the prices p1t , p2t and quantities z1t , z 2t for two intermediate inputs for five periods, t = 1, . . . , 5. It can be seen that the price of the first primary input (labor) is slowly trending upward, while the price of the second primary input (capital services) is trending downward at a faster rate. The quantity of the first primary input is slowly trending downward, while the quantity of the second primary input is trending upward at a rapid rate. Thus normal substitution effects are taking place over time. The price of the first intermediate input (general imports) is trending downward, while the quantity trends upward. Finally, the price of the second intermediate input (energy imports) is rising rapidly, while the corresponding quantity falls. The trends in these prices and quantities are fairly smooth.

(

)

(

(

)

)

(

)

Table 7.1 Primary and Intermediate Input Prices and Quantities Period t

w1t

w 2t

p1t

p2t

x 1t

x 2t

z1t

z 2t

1

1.00

1.00

1.00

1.00

60

40

50

50

2

1.02

0.95

0.99

1.10

59

42

52

48

3

1.04

0.90

0.95

1.20

58

50

55

45

4

1.05

0.82

0.92

1.30

57

57

57

43

5

1.06

0.75

0.90

1.40

56

65

60

40

Productivity Measurement in the Public Sector 249 The most commonly used index number formulae that are used to aggregate prices and quantities are the Laspeyres, Paasche, Fisher, and Törnqvist price indexes . Once the sequence of aggregate prices has been formed using these formulae, the corresponding quantity indexes are formed by dividing the period t value for the aggregate by the corresponding price indexes. Let wt and xt denote the price and quantity vectors for primary inputs for t = 1, . . . , 5. Then the period t fixed base Laspeyres, Paasche, Fisher, and Törnqvist price indexes are defined as follows for t = 1, . . . , 5:26

( ) P ≡ P (w , w , x , x ) ≡ w ⋅ x / w ⋅ x ; (7.6) P ≡ P (w , w , x , x ) ≡ P (w , w , x , x ) P (w , w , x , x ) ; (7.7) P ≡ P (w , w , x , x ) ≡ exp ∑ (1 / 2) (s + s ) ln (w / w ) (7.8) PL t ≡ PL w1 , w t , x1 , x t ≡ w t ⋅ x1 / w1 ⋅ x1 ; 27 (7.5)

P

F

T

t

t

1

T

t

1

F

t

t

t

1

L

t

1

t

1

P

1

n =1

t

1

t

1

t

t

N

t

t

1

F

1 n

n

t

1

t

t

1

n

t

1/ 2

1 n

where snt ≡ wnt xnt / w t ⋅ x t is the nth primary-input cost share in period t. The four fixed-base primary-input price indexes defined by (7.5)–(7.8) are listed in columns 2–5 of Table 7.2, using the primary-input data listed in Table 7.1. These period t fixed-base indexes are denoted by PLX t , PPX t , PFX t and PTX t . It can be seen that PLX t is always greater than PPX t for t = 2, . . . , 5 and the gap between the fixed-base Laspeyres and Paasche price indexes gradually becomes greater as time marches on. Since the Fisher index PFX t is the geometric mean of PLX t and PPX t , it lies between these two equally plausible fixed-basket- type price indexes. Note that the fixed-base Törnqvist price index, PTX t , is quite close to its fixed-base Fisher counterpart, PTX t .28 An alternative to the use of fixed-base indexes is to use chained indexes. Consider how a chained Laspeyres price index, say PLX t * , is formed. For periods 1 and 2, the chained indexes coincide with their fixed-base counterparts; that is, PLX 1* ≡ 1 and PLX 2* ≡ PLX 2 = PL w1 , w 2 , x1 , x 2 . The period 3 chained Laspeyres price index is defined as the period 2 chained index level, PLX 2*, multiplied by PL w 2 , w 3 , x 2 , x 3 , which is (one plus) the Laspeyres rate of change of input prices going from period 2 to period 3. In general, the chained Laspeyres price level in period t+1 is equal to the corresponding

(

)

(

)

Table 7.2 Fixed Base and Chained Laspeyres, Paasche, Fisher, and Törnqvist Primary-Input Price Indexes Period t PLXt

PPXt

PFXt

PTXt

PLXt*

PPXt*

PFXt*

PTXt*

1

1.00000

1.00000

1.00000

1.00000

1.00000

1.00000

1.00000

1.00000

2

0.99200

0.99089

0.99145

0.99145

0.99200

0.99089

0.99145

0.99145

3

0.98400

0.97519

0.97958

0.97963

0.98288

0.97844

0.98066

0.98067

4

0.95800

0.93500

0.94643

0.94661

0.95096

0.94314

0.94704

0.94706

5

0.93600

0.89347

0.91449

0.91492

0.92045

0.90957

0.91499

0.91502

250 W. Erwin Diewert Laspeyres price level in period t times the Laspeyres chain link going from period t to t+1; that is, PLX t +1* ≡ PLX t * PL w t , w t +1 , x t , x t +1 . The chained Paasche, Fisher, and Törnqvist input price indexes, PPX t * , PFX t * and PTX t * are formed in a similar fashion, except that the Paasche, Fisher, and Törnqvist chain link formulae PP w t , w t +1 , x t , x t +1 , PF w t , w t +1 , x t , x t +1 , and PT w t , w t +1 , x t , x t +1 are used, instead of the Laspeyres chain link formula PL w t , w t +1 , x t , x t +1 . The chained Laspeyres, Paasche, Fisher, and Törnqvist input price indexes, PLX t * , PPX t * , PFX t * and PTX t * , are listed in the last four columns of Table 7.2. It can be seen that the use of the chained indexes dramatically reduces the spread between these four types of indexes as compared to their fixed-base counterparts. In period 5, the percentage difference between the Laspeyres and Paasche fixed-base indexes is 4.8%, but the difference between the chained Laspeyres and Paasche indexes is only 1.6%. In period 5, the percentage difference between the superlative Fisher and Törnqvist fixed-base indexes is 0.047%, but the difference between the chained Fisher and Törnqvist indexes is only 0.003%, which is negligible. Thus chaining has substantially reduced the spread between the four most popular index number formulae that are used in empirical applications. This will generally happen if annual data are used so that trends in prices and quantities are fairly smooth.29 The period t primary input implicit quantity indexes that correspond to the input price indexes listed in Table 7.2 can be obtained by dividing the period t aggregate input value by the corresponding period t price index.30 The resulting implicit quantity indexes are listed in Table 7.3. Exactly the same methodology can be used to form eight alternative price indexes for intermediate inputs using the data listed in Table 7.1. Denote the Laspeyres, Paasche, Fisher, and Törnqvist intermediate input price indexes for period t by PLZ t , PPZ t , PFZ t and PTZ t and their chained counterparts by PLZ t *, PPZ t *, PFZ t * and PTZ t *, respectively. These indexes are listed in Table 7.4. Viewing the entries in Table 7.4, it can be seen that the fixed-base Laspeyres and Paasche intermediate input price indexes differ by almost 5% in period 5, while the Fisher and Törnqvist fixed-base indexes are virtually the same. The chained Laspeyres and Paasche intermediate input price indexes differ by about 1% in period 5, while the

(

(

(

)

(

)

(

)

)

)

Table 7.3 Fixed-Base and Chained Implicit Laspeyres, Paasche, Fisher, and Törnqvist Primary-Input Quantity Indexes Period t QLXt

t QPX

QFXt

QTXt

QLXt*

t* QPX

QFXt*

QTXt*

1

100.000

100.000

100.000

100.000

100.000

100.000

100.000

100.000

2

100.887

101.000

100.944

100.943

100.887

101.000

100.944

100.943

3

107.033

108.000

107.515

107.510

107.154

107.640

107.397

107.396

4

111.263

114.000

112.623

112.602

112.086

113.016

112.550

112.548

5

115.502

121.000

118.219

118.164

117.453

118.859

118.154

118.150

Productivity Measurement in the Public Sector 251 Table 7.4 Fixed-Base and Chained Laspeyres, Paasche, Fisher, and Törnqvist Intermediate Input Price Indexes Period t

PLZt

PPZt

PFZt

PTZt

PLZt*

PPZt*

PFZt *

PTZt*

1

1.00000

1.00000

1.00000

1.00000

1.00000

1.00000

1.00000

1.00000

2

1.04500

1.04280

1.04390

1.04390

1.04500

1.04280

1.04390

1.04390

3

1.07500

1.06250

1.06873

1.06874

1.07226

1.06587

1.06906

1.06906

4

1.11000

1.08340

1.09662

1.09664

1.10102

1.09198

1.09649

1.09649

5

1.15000

1.10000

1.12472

1.12475

1.13313

1.12050

1.12680

1.12680

Table 7.5 Fixed-Base and Chained Implicit Laspeyres, Paasche, Fisher, and Törnqvist Aggregate Input Price Indexes Period t

PLYt

PPYt

PFYt

PTYt

PLYt*

PPYt*

PFYt*

PTYt*

1

1.00000

1.00000

1.00000

1.00000

1.00000

1.00000

1.00000

1.00000

2

1.01850

1.01672

1.01761

1.01761

1.01850

1.01672

1.01761

1.01761

3

1.02950

1.01716

1.02331

1.02331

1.02747

1.02135

1.02441

1.02440

4

1.03400

1.00435

1.01907

1.01919

1.02470

1.01474

1.01971

1.01972

5

1.04300

0.98692

1.01457

1.01490

1.02346

1.00923

1.01632

1.01635

corresponding chained Fisher and Törnqvist indexes are exactly the same. Thus again, chaining reduces the spread between indexes for this data set. Use the same methodology to form eight alternative aggregate input price indexes for both primary and intermediate inputs using the data listed in Table 7.1. Denote the Laspeyres, Paasche, Fisher, and Törnqvist aggregate input price indexes for period t by PLY t , PPY t , PFY t and PTY t and their chained counterparts by PLY t * , PPY t *, PFY t * and PTY t * , respectively. These indexes are listed in Table 7.5. Viewing the entries in Table 7.5, it can be seen that the fixed-base Laspeyres and Paasche aggregate input price indexes differ by approximately 5% in period 5, while the Fisher and Törnqvist fixed-base indexes differ by 0.0003. The chained Laspeyres and Paasche intermediate input price indexes differ by about 2% in period 5, while the corresponding chained Fisher and Törnqvist indexes differ by 0.00003. Thus again, chaining reduces the spread between indexes. Also, the superlative indexes (whether they are fixed base or chained) are much closer to each other than are the corresponding Laspeyres and Paasche indexes.31 Once the aggregate input price indexes have been calculated, corresponding aggregate implicit input quantity or volume measures can be calculated by deflating each

252 W. Erwin Diewert period’s total input costs by the appropriate aggregate input price index. Thus the period t fixed-base implicit aggregate input quantity that corresponds to fixed-base Laspeyres price aggregation, QLY t , is defined as follows for t = 1, . . . , 5:32 QLY t ≡ (w t ⋅ x t + pt ⋅ z t )/PLY t . (7.9)

The other seven aggregate implicit input quantity levels, QPY t, QFY t, QTY t, QlY t *, QPY t *, QFY t *, QTY t *, are defined in an analogous manner. The eight aggregate implicit input quantity indexes are listed in Table 7.6. Note that the value of gross output produced during period t, say VY t, is equal to the value of aggregate input used during period t; that is, we have for t = 1, . . . , 5: VY t ≡ w t ⋅ x t + pt ⋅ z t = PLY t QLY t = PPY t QPY t = … = PTY t *QTY t * (7.10)

Thus for each period t, the price of gross output times the corresponding quantity of gross output is equal to the nominal value of aggregate input for each of our eight pairs of gross output price and quantity indexes. The differences between the Laspeyres and Paasche fixed-base and chained aggregate implicit input indexes grow over time and are fairly substantial by period 5, but the other four superlative indexes are reasonably close to each other. With no actual output prices and quantities available for our imaginary public- sector production unit, we set the aggregate output price and quantity levels for period t to equal the corresponding aggregate input price and quantity levels that are listed in Tables 7.5 and 7.6. Thus if we choose to use the fixed-base Laspeyres formula to form aggregate input prices, the period t output price is set equal to PLY t listed in Table 7.5, and the period t aggregate output quantity is set equal to QLY t listed in Table 7.6. The period t gross output TFP level using the fixed-base Laspeyres price index formula is defined as TFPLY t , equal to period t gross output using Laspeyres fixed-base price indexes divided by the period t aggregate input level using fixed base Laspeyres price indexes, which leads to the result TFPLY t ≡ QLY t / QLY t = 1 for each t. Using the other seven methods for aggregating inputs similarly leads to gross output productivity levels

Table 7.6 Fixed-Base and Chained Laspeyres, Paasche, Fisher, and Törnqvist Aggregate Implicit Input Quantity Indexes Period t

QLYt

t QPY

QFYt

QTYt

QLYt*

t* QPY

QFYt*

QTYt*

1

200.000

200.000

200.000

200.000

200.000

200.000

200.000

200.000

2

200.648

201.000

200.824

200.824

200.648

201.000

200.824

200.824

3

205.508

208.000

206.750

206.750

205.913

207.147

206.529

206.530

4

207.863

214.000

210.909

210.883

209.749

211.808

210.776

210.773

5

209.118

221.000

214.977

214.907

213.110

216.114

214.607

214.602

Productivity Measurement in the Public Sector 253 that are identically unity; that is, TFPPY t ≡ QPY t /QPY t = 1, TFPFY t ≡ QFY t /QFY t = 1, . . . , TFPTY t * ≡ QTY t * /QTY t * = 1 for all t. Instead of calculating gross output TFP levels, it is also useful to construct value- added TFP levels.33 Nominal value added for a production unit for period t is defined as the value of outputs produced during period t less the value of intermediate inputs used by the production unit during period t, where an intermediate input is an input that was produced by another domestic or foreign production unit. Thus for our numerical example, period t nominal value added VO t is defined as follows for t = 1, . . . , 5:

VO t ≡ VY t − pt ⋅ z t

(

t

t

t

t

)

(7.11) t

= w ⋅x + p ⋅z − p ⋅z t

= w ⋅x

t

t

where the second equality follows using (7.10). Thus period t nominal value added for our public-sector production unit, VO t , is equal to period t value of primary inputs used in the unit, wt × xt, for each period t. Thus period t nominal value added VO t is unambiguously defined. But how exactly should the price of real value added and the corresponding quantity or volume be defined? We will follow the methodology that was used in the Producer Price Index Manual34 and simply apply normal index number theory to the components of value added, treating all prices and output quantities as positive numbers, but changing the sign of intermediate input quantities from positive to negative.35 Thus the fixed-base Laspeyres, Paasche, Fisher, and Törnqvist price indexes for period t value added are defined as follows for t = 1, . . . , 5:

PLO t ≡ PLY t QLY 1 − pt ⋅ z 1 / PLY 1QLY 1 − p1 ⋅ z 1 ; (7.12) PPO t ≡ PPY t QPY t − pt ⋅ z t / PPY 1QPY t − p1 ⋅ z t ; (7.13) 1/ 2

PFO t ≡ PLO t PPO t ; (7.14) (½ ) sT 11 + sT 1t ln PTY t / PTY 1 + (½ ) sT 21 + sT 2t PTO t ≡ exp (7.15) 1 1 1 t t t × ln p1 / p1 + (½ ) sT 3 + sT 3 ln p2 / p2

( (

)

) (

(

)

) (

(

)

)

where sT 1t ≡ VY t /VO t , sT 2t ≡ − p1t z1t /VO t and sT 3t ≡ − p2t z 2t /VO t for t = 1, . . . , 5. Note that sT 1t > 0, sT 2t < 0, sT 3t < 0, but these value-added “shares” sum up to one; that is, sT 1t + sT 2t + sT 3t = 1 for t = 1, . . . , 5. The value-added price indexes defined by (7.12)–(7.15) are listed in Table 7.7. The fixed-base price index formulae defined by (7.12)–(7.15) can be modified to provide the corresponding chain link price indexes, and these links can be chained together to defined the corresponding Laspeyres, Paasche, Fisher, and Törnqvist value-added chained indexes PLO t *, PPO t *, PFO t * and PTO t *. These chained indexes are also listed in Table 7.7. As usual, the Laspeyres type price indexes are higher (after period 1) than the corresponding Paasche type indexes, and the spread between the fixed-base Laspeyres and

254 W. Erwin Diewert Table 7.7 Fixed Base and Chained Laspeyres, Paasche, Fisher, and Törnqvist Value-Added Output Price Indexes Period t

PLOt

PPOt

PFOt

1

1.00000

1.00000

2

0.99200

3

0.98400

4 5

PTOt

PLOt*

PPOt*

PFOt*

PTOt*

1.00000 1.00000

1.00000

1.00000

1.00000

1.00000

0.99089

0.99142 0.99144

0.99200

0.99089

0.99142

0.99144

0.97519

0.97909 0.97963

0.98288

0.97844

0.98052

0.98065

0.95800

0.93500

0.94445 0.94662

0.95096

0.94314

0.94681

0.94705

0.93600

0.89347

0.90920 0.91494

0.92045

0.90957

0.91465

0.91500

Table 7.8 Fixed-Base and Chained Implicit Laspeyres, Paasche, Fisher, and Törnqvist Indexes of Real Value Added Period t

t QLO

t QPO

t QFO

QTOt

t* QLO

t* QPO

t* QFO

QTOt*

1

100.000

100.000

100.000

100.000

100.000

100.000

100.000

100.000

2

100.887

101.000

100.946

100.944

100.887

101.000

100.946

100.944

3

107.033

108.000

107.570

107.510

107.154

107.640

107.412

107.398

4

111.263

114.000

112.859

112.601

112.086

113.016

112.578

112.550

5

115.502

121.000

118.907

118.161

117.453

118.859

118.199

118.153

Paasche price indexes is larger than the spread between their chained counterparts. As usual, the superlative indexes are generally close to each other. Note also that the value- added price indexes listed in Table 7.7 end up below unity in period 5, whereas the gross output price indexes listed in Table 7.5 end up above unity in period 5 (except for the fixed-base Paasche index). The period t implicit real value-added output quantities or volumes that correspond to the four fixed-base value-added price indexes defined by (7.12)–(7.15) are defined by (7.16) and the corresponding chained indexes are defined by (7.17) for t = 1, . . . , 5:

QLO t ≡ VO t / PLO t ; QPO t ≡ VO t / PPO t ; QPO t ≡ VO t / PPO t ; QPO t ≡ VO t / PPO t ; (7.16)

QLO t * ≡ VO t / PPO t * ; QPO t * ≡ VO t / PPO t * ; QPO t * ≡ VO t / PPO t * ; QPO t * ≡ VO t / PPO t * . (7.17)

The above implicit quantity indexes of real value added are listed in Table 7.8. The substantial difference between the fixed-base implicit Laspeyres and Paasche indexes of real value added in period 5 is noteworthy. The primary-input quantity indexes listed in Table 7.3, along with the real value- added output indexes listed in Table 7.8, can be used to form TFP indexes by dividing

Productivity Measurement in the Public Sector 255 period t real value added by the corresponding measure of real input. Thus the period t value-added TFP level using the fixed-base Laspeyres price index formula is defined as TFPLO t ≡ QLO t /QLX t for each t. Using the other seven methods for aggregating inputs similarly leads to the following alternative measures of real value-added TFP for period t:TFPPO t ≡ QPO t /QPX t ;TFPFO t ≡ QFO t /QFX t;TFPTO t ≡ QTO t /QTX t; TFPLO t* ≡ Q LO t* /Q LX t; TFPPO t * ≡ QPO t * /QPX t ; TFPFO t * ≡ QFO t * /QFx t ; TFPTO t * ≡ QTO t * /QTX t . These alternative measures of value added productivity are listed in Table 7.9.36 Since there are no independent measures of output for our production unit, it should be the case that all TFP levels for our public-sector production unit should equal unity. Looking at Table 7.9, it can be seen that this result holds whenever Laspeyres or Paasche indexes are used. For the Fisher and Törnqvist methods of aggregation, it can be seen that this result does not hold, but it does hold to a high degree of approximation.37 This has been a rather lengthy discussion on how to construct price and quantity estimates for a public-sector production unit when information on the price and quantity of the unit’s outputs is not available. In the end, we have seen that it does not really matter very much whether we measure the productivity of the unit on the basis of its gross output or on its value-added output: in the first case, the level of TFP is always equal to unity, and in the second case, it is either equal to unity if Laspeyres or Paasche aggregation is used or almost equal to unity if superlative index aggregation is used. Thus it would seem that the choice of index number formula is not very relevant for public-sector production units that do not have independent output measures. While this is true as far as the calculation of TFP is concerned, it is definitely relevant when measuring the size of the real value added (and gross output) of public-sector production units. From Table 7.6, it can be seen that the fixed-base Laspeyres and Paasche estimates of real gross output in period 5 were 209.1 versus 221.0, while from Table 7.8, the fixed-base Laspeyres and Paasche estimates of real value added in period 5 were 115.5 versus 121.0. These are substantial differences: the choice of index number formula does matter.

Table 7.9 Fixed-Base and Chained Laspeyres, Paasche, Fisher, and Törnqvist Indexes of Total Factor Productivity Based on Real Value Added Period t

TFPLOt

TFPPOt

TFPFOt

TFPTOt

TFPLOt*

TFPPOt*

TFPFOt*

TFPTOt*

1

1.00000

1.00000

1.00000

1.00000

1.00000

1.00000

1.00000

1.00000

2

1.00000

1.00000

1.00003

1.00001

1.00000

1.00000

1.00003

1.00001

3

1.00000

1.00000

1.00051

1.00000

1.00000

1.00000

1.00014

1.00002

4

1.00000

1.00000

1.00210

0.99999

1.00000

1.00000

1.00025

1.00002

5

1.00000

1.00000

1.00582

0.99997

1.00000

1.00000

1.00038

1.00002

256 W. Erwin Diewert

7.5. Cost-Based Methods for Valuing Public-Sector Outputs When Information on the Quantities of Non-M arket Outputs Is Available In this section, we assume that there is quantity information on the non-market outputs produced by a public-sector production unit, but there are no corresponding market prices to value the non-market outputs. The main public sectors at the national or state level where this situation arises are the health, education, and social services sectors.38 At the municipal or local level, the same situation arises for provision of some services such as waste disposal, water, and sewage services.39 The provision of road and highway services arises at both levels of government. The System of National Accounts 1993 recommends valuing publicly provided services at their unit costs of production.40 In particular, Chapter 16 in SNA 1993 notes that if we have quantity information on the numbers of various different types of outputs produced by a public-sector production unit, then Laspeyres or Paasche indexes can be calculated using sales as values for market services and unit costs times quantities produced as values for non-market services.41 We will indicate in the following exactly how this can be done. We will consider the following two cases: • Case 1: The production unit produces only one non-market output. • Case 2: The production unit produces many non-market outputs. Case 1 is easy to deal with if we make use of the algebra that was developed in the previous section. Thus define the period t price and quantity vectors for primary inputs by wt and xt and the period t price and quantity vectors for intermediate inputs by pt and zt as before. However, now we have information on the quantity of output produced by the production unit during period t, say qt for t = 1, . . . , 5. In order to apply the algebra that was developed in the previous section, we need to change the units of measurement for the single output so that output in period 1 is equal to input cost in period 1. Thus define the normalized output for the production unit for period t, Qt, for t = 1, . . . , 5 as follows: Q t ≡ p1 ⋅ z 1 ⋅ w1 ⋅ x1 qt /q1 . (7.18) Thus when t = 1, the quantity of output produced Q1 is equal to total input cost in period 1. The corresponding period t cost-based output price Pt is defined as period t total cost divided by period t normalized output; that is, define Pt for t = 1, . . . , 5 as follows:42

P t ≡ ( pt ⋅ z t ⋅ w t ⋅ x t )/Q t . (7.19)

Productivity Measurement in the Public Sector 257 The Qt and Pt defined by (7.18) and (7.19) are now independent estimates for the quantity and price of gross output produced by the public-sector production unit during period t. Thus period t fixed-base Laspeyres and Paasche type gross output TFP for the public-sector production unit can be defined as TFPGOL t ≡ Q t / QLY t and TFPGOP t ≡ Q t / QPY t , respectively. The remaining 6 input indexes listed in Table 7.6 can be used to define analogous gross output TFP indexes. Value-added TFP indexes can also be defined by adapting the algebra used in the previous section. The new fixed-base Laspeyres and Paasche value-added price indexes that are counterparts to definitions (7.12) and (7.13) in the previous section are now defined as follows:

PLO t ≡ P t Q1 − pt ⋅ z 1 PPO t ≡ P t Q t − pt ⋅ z t

P 1Q1 − p1 ⋅ z 1 ; (7.20) P 1Q t − p1 ⋅ z t . (7.21)

Thus Pt defined by (7.19) has replaced PLY t in (7.12), and Pt defined by (7.19) has replaced in (7.13), and Qt defined by (7.18) has replaced QLY t in (7.12) and QPY t in (7.13) in equations (7.20) and (7.21). The remaining six value-added price indexes that are counterparts to the remaining six value-added price indexes that are listed in Table 7.7 can be defined in an analogous manner. Denote these remaining six indexes for period t by PFO t , PTO t , PLO t *, PPO t *, PFO t * and PTO t *. The value of value added in period t, VO t , is still equal to the value of primary input in period t, w t ⋅ x t . Now period t real value added can be defined in eight different ways by deflating VO t by our new eight alternative real value- added price indexes. Thus we obtain counterparts to the eight real value-added quantity indexes that appeared in Table 7.8; that is, we have QLO t ≡ VO t /PLO t; QPO t ≡ VO t /PPO t ; . . . , QTO t * ≡ VO t /PTO t *. Finally, eight alternative measures of TFP that are counterparts to the Table 7.9 measures of TFP can be obtained by dividing the new real value-added output indexes by the corresponding primary-input quantity indexes; that is, we have TFPLO t ≡ QLO t /QLX t; TFPPO t ≡ QPO t /QPX t ; . . . , TFPTO t * ≡ QTO t * /QTX t * . Now that we have independent measures of the quantity of gross non-market outputs produced by the public-sector production unit, it will no longer be the case that TFP (either on a gross output or a value-added output concept) will be identically (or approximately) equal to unity in all periods. The choice of index number formula is now important for the measurement of TFP as well as for the measurement of sectoral real output. We will now consider the second case, where the public-sector production unit produces many outputs. Obviously, if data on the price and quantity for each input that is used to produce each output can be found, then the production activities of the public-sector unit can be decomposed into separate production functions, and the index number treatment that was explained earlier for a single output can be applied to each separate production activity. Typically, it will be difficult to allocate the fixed inputs used by the public-sector establishment to the separate activities that produce each output.43 However, it may be possible to obtain estimates of the fraction of total establishment costs in a period that can be imputed to each production activity. Thus

258 W. Erwin Diewert define the overall period t price and quantity vectors used by the production unit for primary inputs by wt and xt and the period t price and quantity vectors for intermediate inputs by pt and zt as before. Suppose that the unit produces K outputs and information on the quantity of each output produced by the production unit during period t is available, say qk t for k = 1, . . . , K and t = 1, . . . , 5. Suppose, in addition, that f k t > 0 is the fraction of period t total cost that can be attributed to the production of output k during period t. Approximate cost-based period t output prices for the K outputs can now be defined as follows for all t and k:

(

)

Pk t ≡ f k t pt ⋅ z t + w t ⋅ x t / qk t . (7.22)

Thus we will have period t output price and quantity vectors, P t ≡ P1t , …, PK t and qt ≡ q1t , …, qK t , for the production unit, and normal index number theory can be used to form output aggregates,44 which in turn can be matched up with the corresponding input aggregates to form TFP estimates. The problem with this method for the valuation of non-market outputs is that it will generally be difficult to determine the appropriate cost fractions f k t .45 In other words, typically it will be possible to measure establishment outputs and total cost in each period, but it will be difficult to decompose the total cost into cost components that can be allocated to each individual output so that the vector of unit costs can be calculated.

7.6. The Use of Quality-Adjusted Output Weights In this section, we discuss the use of purchaser or recipient weights to aggregate the non-market outputs of a public-sector production unit. The basic idea is easy to explain. Consider a public-sector establishment that is producing K outputs over T periods, say qk t for k = 1, . . . , K and t = 1, . . . , T. Let qt ≡ q1t , …, qK t be the period t vector of non-market outputs. We assume that these output quantities can be observed. Define the period t price and quantity vectors of intermediate and primary inputs by the usual pr, wt for prices, and zt and xt for quantities. The aggregation of inputs proceeds as was explained in section 7.4. The problem is how exactly can we construct a period t aggregate output price, say Pt and the corresponding aggregate gross output quantity, say Qt? A possible solution to the problem is to use a vector of user-based relative valuations for the K outputs.46 Let ωk > 0 represent the relative value to users or recipients of output k for k = 1, . . . , K and let ω ≡ [ω1 , , ω K ] be the vector of weights. These weights can be regarded as quality adjustment factors: the higher the weight, the more recipients or users of the non-market outputs of the establishment value the particular output.47 The period t output aggregate can be defined as the weighted sum of the individual period to output quantities using the vector ω as weights, and the corresponding period t

Productivity Measurement in the Public Sector 259 non-market aggregate price P t * can be defined as period t total cost divided by Q t *; that is, we have the following definitions for t = 1, . . . , T:48

Q t * ≡ ω ⋅ qt ; (7.23)

P t * ≡ pt ⋅ z t + w t ⋅ x t / Q t * . (7.24) Now define the aggregate normalized period t output price and quantity by P t ≡ P t * /P 1* and Q t ≡ P 1*Q t * and we can apply the algebra that was developed for Case 1 in section 7.5 to the normalized output price and quantity that we have just defined.49 It should be noted that this method can in principle deal with the introduction of new non-market goods and services; all that is required is a valuation weight for a new commodity relative to the weights for continuing commodities.50 The advantage of this welfare weights method for valuing the non-market outputs of a public-sector production unit over the section 7.5 method is that the present method does not require estimates for the unit cost of production for each non-market output. Of course, the problem with the present method for valuing non-market outputs is that it will be difficult to determine the appropriate vector of output weights ω.51 If experts cannot agree on the appropriate weights, this puts statistical agencies in a difficult position since their estimates of output and input should be objective and reproducible. In the following two sections, we will consider examples of how the use of output valuation weights could work in practice.

7.7. Quality Change, Unit Values, and Linking Bias Technical progress occurs in the public sector, just as it occurs in the private sector. When a new product appears in the private sector during a time period, statistical agencies that construct price indexes face a problem: there is no price in the previous period that can be matched to the new product. Hence, historically, statistical agencies have ignored the existence of the new product during the period when it first appears, but in the second (or later) period of its existence, the new product can be treated in the normal way in which a Laspeyres, Paasche, or other index going from one period to the next is constructed because price and quantity information on the new product is now available for two consecutive periods. Thus the new product is linked in to an existing index that excluded the new product. The problem with this procedure is that it can lead to biased price and quantity indexes.52 We will illustrate the problem by analyzing an artificial example that is due to Schreyer (2010, 21). Schreyer supposes that there is a clinic that offers treatments for eye surgery. In period 1, only a traditional treatment exists. In period 2, a laser surgery alternative is introduced

260 W. Erwin Diewert that is equivalent to the traditional treatment but has a lower unit cost. His data cover three periods. Let q1t and q2t denote the number of traditional and laser treatments done in period t. Schreyer assumes that the period t total costs for performing the traditional and laser surgeries are C1t and C2t , respectively. Thus period t unit costs for the two types of treatment, c1t and c2t , are defined as ck t ≡ Ck t / qk t for k = 1, 2 and t = 1, 2, 3. These data are listed in Table 7.10. Following the cost-based methodology explained in section 7.5, the normalized unit costs for each sector can be used as prices to value the outputs of each sector. Thus the normalized cost-based output prices for the first sector are defined as P1t ≡ c1t / c11 for t =1, 2, 3 and the (normalized) cost-based output prices for the second sector are defined as P2t ≡ c2t / c22 for t = 2, 3. The normalized quantities for each sector k in period t, Qk t , are defined as the sectoral total costs Ck t divided by the corresponding normalized output price, Pk t, so we have Q1t ≡ C1t /P1t = q1t c1t for t = 1, 2, 3 and Q2t ≡ C2t /P2t = q2t c2t for t = 2, 3. These normalized output prices and quantities are also listed in Table 7.10. Since we are interested in productivity measurement in this survey, we will augment Schreyer’s data by adding some detail on the decomposition of treatment costs into price and quantity components. For simplicity, we suppose that there is only one input that is used in each treatment, and the price of this input in period t for treatment k, wk t , is always equal to unity, so we have wk t = 1 for k = 1, 2 and t = 1, 2, 3. Thus the quantity of input used in treatment k for period k, x k t , is equal to the corresponding total cost Ck t divided by wk t . Thus for k = 1, 2 and t = 1, 2, 3, we have: x k t ≡ Ck t /wk t = Ck t . (7.25)

The input prices wk t and the corresponding input quantities x k t = Ck t are also listed in Table 7.10. Period t total cost, Ct, is defined as the sum of the two treatment total costs; that is, we have C t ≡ C1t + C2t for t = 1, 2, 3. Since the sectoral input prices wk t are always equal to unity, it can be seen that any reasonable index number estimate for an aggregate input price will be equal to unity as well. Thus let the period t aggregate input price be defined as W t ≡ w1t = w2t = 1 for t = 1, 2, 3. Define period t aggregate input Xt as total cost Ct divided by the period t input price index Wt, so we have X t ≡ C t /W t = x1t + x2t for t = 1, 2, 3. Ct, Wt, and Xt are listed in Table 7.11.

Table 7.10 Data for Schreyer’s Laser Surgery Example t

q1t

q2t

w1t

w 2t

( ) ( )

C 1t x 1t C 2t xt2

c1t

c2t

P1t

P2t

Q1t

Q2t

1

50

0

1.0

1.0

5000

0

100

__

1.0

__

5000

0

2

40

10

1.0

1.0

4000

900

100

90

1.0

1.0

4000

900

3

5

45

1.0

1.0

500

4050

100

90

1.0

1.0

500

4050

Productivity Measurement in the Public Sector 261 From Table 7.10, it can be seen that the normalized cost-based output prices for both treatments are equal to one for all periods, except that the output price for laser treatments in period 1, P21, is missing. How should an aggregate output price be constructed, given that there is a missing price for the second commodity in period 1? Following standard statistical agency practice in past years, it is natural to use the price movements in the first commodity as the deflator for the value of both outputs in period 2. Letting Pt be the aggregate output price index, this methodology leads to the following definitions for Pt for t = 1, 2: P 1 ≡ P11 = 1; P 2 ≡ P12 = 1. The output price index that takes us from period 2 to period 3 could be the Laspeyres, Paasche, Fisher, or Törnqvist index, but since both output prices remain constant over these two periods, all of these bilateral price indexes will remain constant as well. Thus the aggregate output price index will remain constant over periods 2 and 3. Putting this all together, we will have Pt = 1 for t = 1, 2, 3. The corresponding period t aggregate quantity, Qt, is defined as aggregate cost Ct divided by Pt and thus Q t ≡ C t /P t = C t for t = 1, 2, 3. TFPt is defined as aggregate output Qt divided by aggregate input Xt; that is, TFP t ≡ Q t /X t = 1 for t = 1, 2, 3 since both Qt and Xt turn out to equal aggregate period cost, Ct, for all t. Thus using the usual linking methodology to deal with new products, we end up showing that the introduction of a new more productive technology has led to no measured productivity gains. Ct, Wt, Xt, Pt, Qt, and TFPt are all listed in Table 7.11. Since each treatment gives equivalent results to recipients of the treatment, an alternative measure of aggregate clinic output can be obtained by simply adding up the number of treatments. Thus we define a utility-based measure of output in period t, Qu t * , by performing this addition and define the corresponding unit value price, Pu t *, by deflating total cost Ct by Qu t * . Thus for t = 1, 2, 3, we have: Qu t * ≡ q1t + q2t ; Pu t * ≡ C t /Qu t * . (7.26)

In keeping with our convention that the aggregate output and input prices should equal unity in the base period, we normalize the price series Pu t * by dividing each price by the price in the base period Pu t * and the quantity series Qu t * is normalized by multiplying each quantity by Pu t *. Denote the resulting normalized aggregate output price and quantity series for period t by Pu t and Qu t . For t = 1, 2, 3, we have the following definitions:

Table 7.11 Alternative Measures of Aggregate Output, Input, and TFP for the Schreyer Example t

Ct

Wt

Xt

Pt

Qt

TFP t

Qut*

Put *

Qut

Put

TFPut

1

5000

1.0

5000

1.0

5000

1.000

50

100

5000

1.00

1.000

2

4900

1.0

4900

1.0

4900

1.000

50

98

5000

0.98

1.020

3

4550

1.0

4550

1.0

4550

1.000

50

91

5000

0.91

1.099

262 W. Erwin Diewert

Pu t ≡ Pu t * / Pu1* ; Qu t ≡ Qu t * Pu1* = Qu t * C1 / Qu1* .

(7.27)

The utility-based output measure, Qu t , can now be used to define a utility-based measure of TFP that is equal to Qu t divided by our measure of aggregate input Xt; that is, define TFPu t ≡ Qu t /X t for t = 1, 2, 3. The series Pu t *, Qu t * , Pu t , Qu t and TFPu t are listed in Table 7.11. Comparing TFPu t with our earlier measure TFPt, it can be seen that our new measure shows that productivity increased substantially during periods 2 and 3, as compared to our old measure that showed no increase in productivity.53 An example of how linking bias can lead to estimates of output that have a downward bias occurred when Griliches and Cockburn (1994) discussed how statistical agencies treated the introduction of generic drugs into the marketplace. A generic drug has the same molecular composition as the corresponding brand name drug and so instead of treating the generic and brand name drug as separate products and linking in the generic to the price index for drugs when the generic drug first appears on the marketplace, it may be preferable to treat the products as being equivalent, which will lead to a higher aggregate output of drugs, as in the preceding example.54 Recall that the utility-oriented approach to output measurement initially measured period t aggregate output as Qu t * = q1t + q2t and our final output measure Qu t was proportional to Qu t * . Thus it can be seen that Schreyer’s model is a special case of the more general quality-adjustment model that was defined by equation (7.23) in the previous section. This equation is Q t * ≡ ω ⋅ qt, which in turn is equal to ω1q1t + w2 q2t when there are only two outputs. Thus if we set ω1 = ω2 = 1, the general quality-adjustment model in the previous section reduces to the method used by Schreyer. For Schreyer’s example, it was easy to determine the weights ω1 and ω2. In most real-life examples, it will be more difficult to determine the appropriate weights. There is an alternative method for dealing with new goods or services that is due to Hicks (1940, 114). In the period before the new commodity appears, we could imagine (or estimate)55 an imputed price for the new product that would just cause potential purchasers to demand zero units of it. Now match up this imputed price with the corresponding quantity (which is 0) in the period that precedes the introduction of the new good and apply normal index number theory. For Schreyer’s example, since the two commodities are close to being identical, an imputed price for the laser treatment that is slightly higher than the actual period 1 price for the traditional treatment would be appropriate. From Table 7.10, we see that the price for the traditional treatment in period 1 is c11 = 100. Thus, following Hicks’s methodology, set the imputed price for the laser treatment, c21, equal to 100 as well. Now go to Table 7.10 and use the output quantity data that is in the q1t and q2t columns and the corresponding price data that are in the unit cost columns c1t and c2t but replace the missing price for c11 by 100. Apply normal index number theory to this new set of price and quantity data. Calculate the fixed base Laspeyres, Paasche, Fisher, and Törnqvist price indexes for each period t, PL t , PP t , PF t , PT t , and the corresponding quantity indexes, QL t , QP t , QF t , QT t , which are defined as total

Productivity Measurement in the Public Sector 263 period t cost Ct divided by the corresponding price index for period t. These fixed-base output price and quantity indexes are listed in Table 7.12. As usual, the superlative price indexes, PF t and PT t , are fairly close to each other and hence so are the companion superlative quantity indexes, QF t and QT t . The use of these superlative quantity indexes as the clinic’s output measure would lead to TFP indexes that end up around 1.048 in period 3. The use of the fixed-base Laspeyres formula to aggregate output prices leads to a constant price index (i.e., PL t = 1 for all periods t); the corresponding output quantity index (listed as QL t in Table 7.12) is actually a fixed-base Paasche quantity index and it takes on exactly the same values as Qt in Table 7.11. Thus for this example, the use of the fixed-base Laspeyres formula for aggregating output prices leads to the same (downward-biased) TFP indexes that occurred when we used the linking methodology that was explained in the beginning of this section. What is striking is that the use of the fixed-base Paasche formula to aggregate output prices leads to the output price index listed as PP t in Table 7.12 (the corresponding output quantity index is listed as QL t ), and this price index is exactly equal to the utility weighted output price index Pu t that is also listed in Table 7.11. Thus the productivity index that results from the use of PP t is exactly equal to the utility weights TFP index, TFPu t , that is listed in Table 7.11. The reason why this equality occurs is due to the fact that the output quantity index QPt that corresponds to PPt turns out to be equal to the fixed-weight index QP t ≡ c11q1t + c21q2t = 100q1t + 100q2t = Qu t . The preceding analysis shows that there are at least two methods that can be used to mitigate possible bias due to the introduction of new products into the public sector: the use of utility weighting of outputs, or the use of Hicksian imputed prices to value the new product in the period before its introduction. The first method requires estimates for relative welfare or utility weights for the new product relative to existing products. The second method requires estimates for shadow prices that value the new product relative to existing products (these shadow prices are essentially proportional to utility weights) in the period prior to the introduction of the new product. At first glance, it appears that the second method is preferable, since the use of Hicksian shadow prices is consistent with normal consumer theory. Using a superlative index number formula to aggregate either prices or quantities, the resulting volume or quantity indexes can be consistent with utility-maximizing behavior on the part of users of the products, where the functional Table 7.12 Fixed-Base Laspeyres, Paasche, Fisher, and Törnqvist Price and Quantity Indexes for the Schreyer Data with Hicksian Imputation Period t

PLt

PPt

PFt

PTt

QLt

1

1.00000

1.00000

1.00000

1.00000

5000.000 5000.000 5000.000 5000.000

2

1.00000

0.98000

0.98995

0.99037

4900.000 5000.000 4949.747 4947.642

3

1.00000

0.91000

0.95394

0.95419

4550.000 5000.000 4769.696 4768.436

QPt

QFt

QTt

264 W. Erwin Diewert form for the underlying utility function is reasonably flexible. The use of the first method essentially assumes that purchasers of the products have linear sub-utility functions that remain constant from period to period. However, the apparent superiority of the second method rests on the assumption of utility-maximizing behavior on the part of purchasers and on the existence of market prices for the commodities under consideration. In the case of goods and services supplied by the public sector, the usual justifications for the economic approach to index number theory do not apply, and so it is not clear that the second method for mitigating new good bias is superior to the first method. A final method that might be used to quality adjust the outputs of public-sector production units is the use of hedonic regression methods.56 A hedonic regression model regresses the price of a product on the price-determining characteristics of the product. Once a hedonic regression has been determined, the relative value of a new product with certain characteristics can be determined relative to existing products using a hedonic regression model. Thus if a new public-sector output has a mix of characteristics that is similar to products with similar characteristics that sell in the marketplace, then the value of the new public-sector output relative to existing marketplace outputs that are similar could be determined using a hedonic regression that involves only marketed products. Thus a hedonic regression model may provide a scientific method for determining the valuation weights ωk that made their appearance in the previous section. The problem with this suggestion is that the characteristics of the public-sector outputs may be quite different from characteristics of “similar” products that appear in the market sector; that is, there may be no such similar products. However, in some situations, such as the valuation of subsidized housing, the hedonic regression methodology may well work in a satisfactory manner.

7.8. Specific Measurement Issues In this section, a few of the measurement issues that arise in measuring outputs in specific subsectors of the public sector will be discussed. The focus will be on possible methods for choosing the utility-oriented valuation weights ωk that made their appearance in section 7.6. For a comprehensive discussion on how to measure outputs for the entire public sector, see Hill (1975) and Atkinson (2005). For a detailed discussion on how to measure education and health outputs, see Schreyer (2010, 2012a).

7.8.1. The Education Sector Hill (1975, 48) recommended that output in the public education sector should be measured by pupil hours of instruction with possible quality adjustment for the number of students in the classes of instruction under consideration.57 Hill (1975, 46) did not favor any quality adjustment for class failure rates or for class performance on test scores. He argued that the output of private driving classes is measured by fees collected and hence

Productivity Measurement in the Public Sector 265 is proportional to the number of students taking the driving course of instruction, and he observed that there is no quality adjustment of the outputs of driving schools for subsequent failures when students take their driving tests. Hence, by analogy, there should be no adjustment for failures when students fail their classes. This argument is not convincing since it may be more reasonable to quality adjust the private-sector driving school output for student failure of the subsequent driving tests. Atkinson (2005, 128) noted that the United Kingdom uses the number of full-time equivalent students as the output measure in the national accounts.58 Atkinson (2005, 130) basically endorsed this method for measuring school outputs but noted that there should be a switch from registered pupil numbers to actual school attendance numbers, and for pupils aged 16 and over, some account of school examination success should be taken into account. Schreyer (2010, 37) recommended (as a first step) that education output for primary and secondary education services be measured by pupil hours, differentiated by the level of education and possibly other characteristics. However, Schreyer (2010, 42) later indicated how pupil hours could be quality adjusted by average test results for the class under consideration: The target measure for the quality-adjusted volume change of education services is the change in the number of pupil hours (H) multiplied by the quality of teaching. The indicator for the quality of teaching is average scores (S) divided by the change of pupil hours per pupil (H/N). Division by H/N is necessary because pupil attainments are influenced by possible changes in the number of lessons and this influence should be eliminated to arrive at quality of one pupil hour.

Schreyer (2010, 42) went on to show that the change in the volume of educational services going from period t–1 to t for the class under consideration was equal to the following expression:

Change in volume = H t /H t −1 St /St −1 t

t

t −1 t −1

= N S /N S .

{H /N t

t

}

H t −1 /N t −1 (7.28)

Thus the final Schreyer measure of output in period t for the class under consideration is proportional to NtSt, the number of students in the class, Nt, times the average class score, St, for a test that appropriately measures what has been learned in the class. Thus if there are K classes in scope for the educational output index where the number of pupils in class k in period t is N k t and the average test score for students in the class is Sk t, then period t output is proportional to Q t = Sk =1K Sk t N k t . This output measure fits in with the general class of quality-adjusted output measures that were discussed in section 7.6, where Qt was proportional to ω ⋅ qt ≡ ∑ k =1K ω k qk t or, more generally, proportional to ωt ⋅ qt ≡ ∑ k =1K ω k t qk t , where the vector of quality-adjustment factors ωt can change as t changes. Thus the average test score Sk t plays the role of a period t quality-adjustment factor ω k t and the number of students in class k during period t, N k t , plays the role of an output measure qk t that is not quality adjusted.

266 W. Erwin Diewert The Schreyer volume measure is a reasonable one, but it suffers from two defects: • The tests administered in each period may not adequately reflect the actual average increase in knowledge that students have acquired going from period t–1 to t; • The previously mentioned average test results do not take into account the capabilities of the class being tested. The first problem is always a potential problem with using standardized tests to measure the increased capabilities of a class, and so caution must be used in allowing test results to be the dominant factor in measuring educational outputs.59 The second problem could potentially be addressed by testing students at the beginning of each class term and at the end of the class term. Suppose there are K classes in scope during a number of periods indexed by t. Suppose that the number of students in class k during period t is Nkt for k = 1, . . . , K. Suppose further that students are tested at the beginning of period t and at the end of period t on the materials to be covered in the class and that the beginning and end of period test scores for student n in class k during year t are Skntb and Sknte for k = 1, . . . , K and n = 1, …, N k t ≡ N (k, t ). Define the beginning and end of year average test scores, Sk •tb and Sk •te , for class k in year t as follows:

N k ,t N k ,t Sk •tb ≡ ∑ n =1 ( ) Skntb /N k t ; Sk •te ≡ ∑ n =1 ( ) Sknte /N k t . (7.29)

If we attempt to measure educational output by the increase in a student’s knowledge and capabilities due to classroom teaching, then the quality-adjusted output of class k in year t, Qk t *, could be measured as being proportional to the class sum of end of period t test scores less the corresponding sum of beginning of period t test scores; i.e., we have for k = 1, ..., K:

Qk t * ≡ Sn =1N (k ,t ) Sknte − ∑ n =1N (k ,t ) Skntb (7.30) = Sk •te N k t − Sk •tb N k t = Sk •te − Sk •tb N k t

Now set ω k t ≡ Sk •te − Sk •tb and qk t ≡ N k t and define (preliminary)60 total period t output as Q t * ≡ ∑ k =1K ω k t qk t . Thus we see that the number of students in class k can play the role of an unadjusted measure of class k output, and the average difference in the class test scores between the end and beginning of the year can play the role of a quality- adjustment factor. It may be very difficult to design tests given at the beginning and end of a class that will accurately measure the effect of the teaching on increasing student knowledge and capabilities. Moreover, because the quality-adjustment factor is a difference, the resulting output volume measures could turn out to be quite volatile.61 Obviously, the issues surrounding the quality adjustment of educational output measures are far from settled.

Productivity Measurement in the Public Sector 267

7.8.2. The Health Sector Hill (1975) has an extensive (and thoughtful) discussion of the problems associated with the measurement of health services in both the private and public sectors. Hill (1975. 33) noted that there are two general approaches to the measurement of health service outputs: As already explained, the appropriate measure of the output of the health industry or branch in the context of economic accounting is the treatment actually provided to consumers or patients. An alternative view is to regard medical treatment as only a means to an end, namely achieving an improvement in health, and to seek to measure the output in these terms.

Thus Hill endorsed output measures that are based on the amount of medical treatment provided to patients (such as patient days in hospitals or number of visits to general practitioners), irrespective of the outcome of such treatments. But if a particular public-sector medical treatment accomplishes nothing, does it make sense to treat the expenditures associated with the treatment as a positive output? Thus the issues are similar to the measurement of outputs in the public education sector. Patient hours (in hospitals) or patient days replace pupil hours or pupil days as unadjusted measures of output in the hospital sector compared to measures of output in the public education sector, and number of patients treated replace number of students in classes as alternative unadjusted measures of output in the health sector as compared to the education sector. These measures do not reflect any improvements (or failures) in capabilities of patients treated or of pupils taught. The quality-adjustment methodology that was suggested at the end of section 7.8.1 could be modified to apply to the public health sector. Instead of a test score at the beginning and end of each class, the health sector counterpart would be some measure of the capabilities of a patient before and after the medical treatment. Thus suppose the medical treatment is a hip joint replacement. Before the operation, medical experts would have to devise a “test” score that rated the capabilities of the patient to perform a range of tasks with the impaired leg on a scale of say 1 to 10, with 10 being completely “normal” and 1 being essentially immobile. After the operation with suitable recovery time, the patient would be graded again for mobility using the same scale. However, mobility is not the only issue: before and after the operation, the patient could have various degrees of pain associated with the condition. Again, the degree of pain before and after the operation could be measured on a scale of 1 to 10, but now we would have to face the issue of how to weight the two scales in an overall scale. Assuming that these “test” design issues could be adequately addressed by medical experts, the algebra surrounding equations (7.29) and (7.30) could be adapted to the medical context.62 Obviously, there would be many difficulties associated with implementing this outcome-based methodology. Thus in the face of all the difficulties associated with implementing an outcomes methodology

268 W. Erwin Diewert for health services, it may be necessary to implement Hill’s preferred methodology and just measure the various services that the health sector provides without attempting to measure outcomes. Hill (1975, 36) provided the following description of the type of hospital services that he would measure: For example, suppose an individual enters a hospital for treatment. In general, such treatment can be decomposed into a number of different elements. The following services may be itemized. (1) The provision of food, accommodation and hotel type services. (2) Nursing care. (3) Medical examinations including diagnostic services such as: (i) laboratory tests; (ii) X-ray examinations; (iii) other forms of examinations such as cardiographs, etc. (4) The provision of drugs and other similar remedial treatment. (5) Various specialist services such as: (i) surgery; (ii) radiotherapy; (iii) physio therapy, etc. The above breakdown is only intended to be illustrative.

Using this breakdown of hospital-associated services, there are still some problems associated with measuring the outputs of each of the preceding five components. The number of patient days would probably suffice for measuring outputs for category (1); the number of nurse days could suffice for (2); the number of “standard” examinations would work for (3); the number of drugs administered for (4); and either the number of “standard” interventions for category (5) or the number of hours of each type of intervention that was administered by the hospital. Of course, working out the allocation of total hospital costs to each of the five types of activity would be difficult. If, instead of following Hill’s service-oriented methodology, we followed an outcome- oriented methodology, then we would no longer measure all of the particular services listed in categories (2)–(5): the focus would be on the number of individual patient treatments for various ailments and the outcomes of the hospitalization process. However, we should still measure the “hotel” services that the hospital provides as separate outputs that should be added to the treatment outcome outputs. These food and accommodation services are substituting for the food and accommodation services that are no longer being consumed at the residences of hospital patients. The provision of hotel services for nursing homes and assisted living arrangements are very important components of the outputs of these subsectors of the public sector. It can be seen that the measurement of public health sector outputs is an extremely difficult problem. In practice, national statistical agencies use only very rough measures of output for their public health sectors. Hill (1975, 42–43) listed how various countries measured the output of their health sectors as of 1975. For example, in the Netherlands, the output of the hospital sector was measured by the number of patient days, while the output of other health services was proportional to an index of employment in the non-hospital health sector. For the United Kingdom, the output of the

Productivity Measurement in the Public Sector 269 hospital sector was also proportional to an index of hospital employment, the output of general practitioners was measured by the number of general practitioners, and the output of most other kinds of health services (including dental services) was measured by the number of treatments. Atkinson (2005, 103–124) provided an extensive review of health output measurement issues in the United Kingdom. Atkinson (2005, 106) noted that prior to 2004, hospital outputs were primarily measured by patient days, while other medical outputs were primarily measured by the number of consultations for specific treatments. After 2004, the number of treatments that were recognized as separate categories was greatly increased and cost weights for each category were constructed. This constitutes a big improvement over the UK output measures that were in place in 1975, as described by Hill in the previous extract. Schreyer (2010, 72–106) has an extensive treatment of the measurement issues surrounding health care, including references to the literature as well as a detailed description of methods used to measure health outputs in a large number of countries. Schreyer adopted a treatment-based definition for the outputs of the health sector as his target output concept.63 In practical terms, using a treatment approach to measuring health sector outputs means that the hospital output measure for the treatment of a narrowly defined medical condition would not use patient days as the output measure, but would simply use the number of patients treated for the condition (assuming that average treatment outcomes remain constant from period to period). For an application of the treatment outcome approach to the problem of making international comparisons of health sector real output, see Koechlin, Konijn, Lorenzoni, and Schreyer (2015).64

7.8.3. The Infrastructure, Distribution, and Public Transportation Sectors The public sector in every country provides a vast network of roads and highways that typically can be used free of charge to transport passengers and goods from place to place. From a utility or demander perspective, the output generated by a given stretch of homogeneous road or highway over a period of time should at least include the passenger miles traveled over the road as well as the ton miles of freight that is shipped over the road during the time period. However, even if a household or firm does not use the road in a given period, they may still value the option or possibility of using the road, and thus the road network itself could also be regarded as a valued output from the demander perspective. From the cost or supplier perspective, building the road initially is definitely a cost-determining output. If the public sector also maintains the road, then measures of the utilization of the road (such as passenger miles and ton miles generated by users of the road over the time period) will also be cost-determining outputs. Note that the network weights and the utilization weights will generally be quite different from the user and supplier perspectives. Taking the cost perspective to the valuation

270 W. Erwin Diewert of public-sector outputs, the supplier cost weights will be based on the initial cost of building the road, and the resulting user costs65 will form the network component of total road cost in period t, and the utilization cost component will be based on the sum of labor and material costs plus the maintenance and capital equipment user costs that pertain to period t. Lawrence and Diewert (2006, 215) noted the similarity in measuring the output of a road system with measuring the output of an electricity distributor: The distributor has the responsibility of providing the “road” and keeping it in good condition but it has little, if any, control over the amount of “traffic” that goes down the road. Consequently they argue it is inappropriate to measure the output of the distributor by a volume of sales or “traffic” type measure. Rather the distributor’s output should be measured by the availability of the infrastructure it has provided and the condition in which it has maintained it—essentially a supply side measure.

Lawrence and Diewert (2006, 215) go on to suggest that a comprehensive output measure for a regulated electricity distributor should consist of three components: throughput, network line capacity, and the number of customers.66 Moreover, they followed the national accounts treatment of public-sector production and valued each of the three output components by their imputed cost of production. Lawrence and Diewert (2006, 215) also suggested that the same methodology that treats throughput and the underlying network as separate outputs could be applied to passenger traffic on government-owned or regulated railways and transit systems, to pipelines, to telecommunication providers, and to natural gas distributors. The point is that many distribution production units have the opportunity to behave in a monopolistic manner and thus they are regulated by the government. The regulators typically force the distribution unit to provide services to all potential customers in an area at regulated prices. Hence these prices are not necessarily the prices that would be generated by unregulated markets. This fact has implications for the economic approach to the measurement of TFP that relies on exact index numbers, which in turn relies on the assumption of competitive behavior in both output and input markets.67 While it is reasonable to assume that a regulated firm behaves competitively on input markets, it is not reasonable to assume that regulated firms behave competitively on output markets. Thus a different index number methodology that relies on the estimation of cost functions or on the estimation of unit costs to value outputs (as was explained in section 7.5) is needed to measure productivity, not only for public-sector production units, but also for regulated production units in the distribution, telecommunications, and transportation sectors.68 Thus the unit-cost-based methodology to the measurement of public-sector production and productivity that was pioneered by Scitovsky (1967), Hill (1975), and Schreyer (2010) has a wider application to the regulated part of the private sector. National post offices that offer nationwide mail delivery at regulated prices are another example of a production unit that produces a network availability output as well as

Productivity Measurement in the Public Sector 271 utilization or throughput outputs. The unit-cost-based methodology for output measurement could also be applied to this sector.

7.9. Conclusion This chapter has covered the three main classes of methods used by national income accountants to construct measures of real output for public-sector establishments that produce non-market outputs. If it is difficult or impossible to construct unambiguous measures for the quantities of non-market outputs produced by a public-sector production unit, then aggregate output is typically set equal to a measure of establishment aggregate input and the resulting TFP estimates will show no productivity gains. However, in section 7.4, we showed that measurement was not completely clear-cut in this situation; that is, we looked at the complications that arise if we want to measure the value added of the production unit versus its gross output, and the consistency between the two measures. We also showed that the choice of index number formula will, in general, make a substantial difference to the resulting measures of gross output or value-added growth of public-sector production unit. In section 7.5, we covered the second general method for constructing aggregate output measures for the non-market outputs of a public-sector production unit. Using this methodology, the prices of non-market outputs are set equal to the unit costs of producing the outputs. Once these imputed prices have been determined, normal index number theory can be applied to construct estimates of aggregate non-market output and of TFP of the public-sector unit. Of course the practical problem with this method is that it will typically be difficult to construct suitable measures of unit cost for the non-market outputs. In section 7.6, we covered the final general method for constructing aggregate output measures for public-sector production units. Using this approach, the non-market outputs of a public-sector establishment are aggregated together by using a vector of weights that reflect the relative value of the non-market outputs to users or recipients of the non-market outputs. The corresponding aggregate non-market output price is determined by dividing total establishment cost (less the value of market outputs) by the welfare-oriented non-market quantity index. The advantage of this method over the section 7.5 method is that it is not necessary to form estimates of the unit costs for the non-market outputs. This disadvantage of the section 7.6 method is that it will generally be difficult to determine the appropriate vector of non-market output weights. In section 7.7, we discussed the problem of quality adjustment of non-market outputs, which boiled down to the problem of finding appropriate vectors of non-market output weights. We also discussed the problems associated with linking in new products. In section 7.8, the problems associated with measuring non-market outputs in the education and health sectors were discussed and how the general measurement methods discussed in sections 7.5 and 7.6 could be applied to these sectors. We also noted that there were measurement problems for finding suitable prices for the outputs of regulated

272 W. Erwin Diewert firms that are entirely analogous to the problems associated with finding reasonable prices for the non-market outputs of public-sector units. Regulated firms are regulated because they have some sort of monopoly power. Typically, the regulatory authorities require the regulated firms to provide uniform levels of service over regions or locations at regulated prices. Thus the quantities produced by a regulated firm are typically not the quantities that a price-taking competitive firm would provide at the prices that regulators set. Hence the same methods that are used to value the outputs of non-market producers should be used to measure aggregate output and TFP of regulated firms; that is, the regulated prices should be replaced by marginal or average unit cost prices. Thus the methods discussed in this chapter have some applicability to the measurement of TFP in regulated industries. In section 7.3, we provided an extensive discussion on how to measure the value of capital services in the public sector. These measurement problems deserve a lot more attention than they have received in the past. Statistical agency measures of the value of capital services in the government sector do not include the imputed interest cost of the fixed capital that is used in this sector, due to national income accounting conventions. This convention has led to a very large downward bias in both the nominal and real GDP of all countries, with the bias being bigger for rich countries that generally have larger public sectors than poorer countries. The only cost associated with capital inputs used in the public sector that is allowed in the international System of National Accounts is depreciation. Thus the user costs of government buildings are vastly understated. In addition to the understatement of the costs associated with the use of structures, there is a further understatement due to the complete neglect of land-user costs for government-owned land. Because land does not depreciate, the costs associated with the land that sits under public schools and hospitals are set equal to zero, which is the ultimate understatement! If a government-owned office building were instead rented from the private sector, the explicit rent would be recognized in the SNA and this explicit rent would include the interest opportunity cost of capital that is tied up in the structure and the land plot that supports the structure.69 The neglect of the land that roads and government-owned railways sit on also will lead to a substantial downward bias in the GDP of the public sector. Finally, we conclude this chapter with some observations on the difficulties associated with determining the “right” prices for valuing public-sector outputs that are allocated to households at very low or zero prices. Suppose that a household has preferences over market goods and services and over non-market goods and services that are provided to it by the public sector. Denote the household consumption vectors of non-market and market commodities by y and z, respectively, and suppose that the households preferences can be represented by the utility function f(y, z). Suppose that the household faces the price vector wt for market goods and services and has “income” It to spend on these commodities. The various levels of government allocate the vector yt of public goods and services to the household in period t. We assume that the vector zt solves the following period t utility-maximization problem for the household:

Productivity Measurement in the Public Sector 273

{( )

}

(

)

max z f y t , z : w t ⋅ z ≤ I t ≡ ut = f y t , z t . (7.31)

(

)

Thus z t = d I t , pt , y t where d is the household’s system of conditional market demand functions for market goods and services, conditional on (i) income spent on market commodities It; and (ii) the price vector for market goods and services pt and the household’s allocation of public goods, yt. The conditional expenditure function e that is generated by the utility function f is defined as follows:

(

)

{

(

}

)

e ut , w t , y t ≡ min z w t ⋅ z ; f y t , z ≥ ut = w t ⋅ z t . (7.32)

Suppose that e(ut, wt, y) is differentiable with respect to the components of y when y = yt and let

(

)

pt ≡ ∇ y e ut , w t , y t (7.33)

denote this vector of partial derivatives. Define the household’s augmented period t income, I t * , as follows: It* ≡ It + pt ⋅ y t . (7.34)

Then under suitable regularity conditions, it can be shown that yt, zt is a solution to the following augmented income utility maximization problem:

{( )

}

(

)

max y , z f y , z : pt ⋅ y + w t ⋅ z ≤ I t * ≡ ut = f y t , z t . (7.35)

It can be seen that pt defined by (7.33) is an appropriate vector of shadow prices that value the components of the public-sector quantity vector yt for this household; that is, the components of pt reflect the value of the public goods vector yt from a household welfare perspective. How could we calculate this vector of shadow prices in practice? It would be necessary to estimate the household’s system of market demand functions, d(It, pt, yt), given a time series of data on It, pt, and yt for the household. We could then use the estimated market demand functions and attempt to recover the underlying utility function, f(y, z), up to a cardinalization, and then the corresponding dual expenditure function e could be recovered, and finally the welfare-oriented prices pt could be calculated. Normal index number theory could be applied at this point.70 But there is a problem with the preceding methodology: we cannot fully recover the preferences of the household using this methodology!71 To show why this is the case, replace the original household utility function f(y, z) by F(y, z) ≡ f(y, z) + g(y) where g(y) is a subutility function that is just defined over the public goods. Now assume that the household solves the following conditional utility maximization problem:

{( )

}

max z F y t , z : w t ⋅ z ≤ I t . (7.36)

274 W. Erwin Diewert It can be seen that the zt = d(It,pt,yt), which was the solution to the original utility maximization problem defined by (7.31), is also a solution to the new utility maximization problem defined by (7.36). This shows that a knowledge of the household’s system of conditional market demand functions is not sufficient to fully reconstruct household preferences over market and non-market goods and services,72 and hence it will be difficult to construct prices for non-market commodities from the welfare perspective.73

Notes 1. W. Erwin Diewert, School of Economics, University of British Columbia, Vancouver, BC, Canada, V6T 1Z1, and the School of Economics, UNSW Sydney 2052, Australia. Email: [email protected]. The author thanks Knox Lovell, Kevin Fox, Paul Schreyer, and Robin Sickles for helpful comments and the SSHRC of Canada and the Australian Research Council (DP150100830) for financial support. None of the above is responsible for any opinions expressed in the chapter. 2. This follows the approach pioneered by Jorgenson and Griliches (1967, 1972). For further developments of the index number approach to measuring TFP growth, see Diewert (1976, 1980, 1983, 1992a, 1992b, 2014), Caves, Christensen, and Diewert (1982a, 1982b), Diewert and Morrison (1986), Kohli (1990), Fox and Kohli (1998), Balk (1998, 2003), Schreyer (2001), Diewert and Nakamura (2003), and Inklaar and Diewert (2016). 3. However, there are some significant problems associated with the measurement of capital services inputs and we will spend some time dealing with these difficulties. 4. For a more detailed discussion of valuation principles based on purchaser versus supplier valuations, see Hill (1975, 19–20), Atkinson (2005, 88), and Schreyer (2012a, 261–266). 5. However, from the perspective of measuring the total factor productivity growth of a public-sector production unit, valuation of unpriced outputs by their marginal or unit costs is best, as we shall see later. 6. During recessions, outputs decrease more than inputs due to the short-run fixity of many inputs. Thus production units may be in the interior of their production possibilities sets at times and so movements from a technically inefficient allocation of resources toward the production frontier will improve TFP. 7. Other methods involve the estimation of cost or production functions, but typically, statistical agencies do not have the resources to undertake the required econometric estimation. Nonparametric productivity measurement methods could also be used, but unless price information is used in addition to quantity information on the inputs used and outputs produced by a production unit, these methods often do not generate reasonable estimates in the time series context that is the focus of this chapter. For the application of nonparametric methods in the cross-sectional context, see the pioneering papers by Charnes, Cooper, and Rhodes (1978) and Charnes and Cooper (1985). 8. Atkinson (2005, 88–90) advocated the use of marginal costs to value outputs in the non- market sector. Hill (1975, 19–21) advocated the use of (average) unit costs to value non- market outputs. 9. In section 8.1, we will discuss various suggested methods for estimating output prices and quantities for this sector. It will be seen that there is no general agreement on these methods.

Productivity Measurement in the Public Sector 275 10. A capital input is an input that contributes to production for more than one accounting period. In practice, national income accountants treat capital inputs that last less than three years as nondurable inputs. 11. For simplicity, we assume the geometric model of depreciation where the one period depreciation rate δ remains constant regardless of what the age of the asset is at the beginning of the period. For more on the geometric model of depreciation, see Jorgenson (1989, 1996). 12. This simple discrete time derivation of a user cost (as the net cost of purchasing the durable good at the beginning of the period and selling the depreciated good at an interest rate discounted price at the end of the accounting period) was developed by Diewert (1974, 504; 1980, 472–473; 1992b, 194). Simplified user cost formulae (the relationship between the rental price of a durable input to its stock price) date back to Babbage (1835, 287) and to Walras (1954, 268–269). The original version of Walras in French was published in 1874. The early industrial engineer Church (1901, 907–909) also developed a simplified user cost formula. 13. It should be noted that the user costs that are anti-discounted to the end of the period are more consistent with commercial accounting conventions than the corresponding user costs that are discounted to the beginning of the period; see Peasnell (1981). 14. Diewert (2005, 2010) and Diewert and Wei (2017) derive user cost formula for more gen eral models of depreciation. In particular, one hoss shay or light-bulb depreciation may be a more appropriate model of depreciation in valuing the contribution of structures and long-lived infrastructure assets. Diewert (2005) also discussed in more detail user costs that are formed by discounting or anti-discounting costs and benefits to either the beginning or end of the accounting period. 15. See, in particular, Jorgenson and Griliches (1967, 1972) and Christensen and Jorgenson (1969). 16. Of course, the problem with using ex ante user costs is that there are many methods that could be used to predict asset inflation rates and these different methods could generate very different user costs. For empirical evidence on this point, see Harper, Berndt, and Wood (1989), Diewert (2005), and Schreyer (2012b). 17. The predicted asset inflation rate was set equal to the ex post geometric average inflation rate for the asset over the past 25 years. 18. While Jorgensonian user costs may not be the best for productivity measurement purposes, they are the “right” user cost concept to use when calculating sectoral ex post rates of return. Moreover, Diewert and Fox (2016a) found that even though many Jorgensonian land user costs turned out to be negative, the resulting rates of TFP growth did not differ much from the corresponding TFP growth rates using ex ante or smoothed asset inflation rates. 19. See also Diewert, Nakamura, and Nakamura (2009). 20. Thus the user cost formula (4) should be modified to ut ≡ (1 + r t )U t = (1 + r t ) − (1 − δ) (1 + i t ) + τt P t = r t − i t + δ (1 + i t ) + τt P t where τt is the period t specific tax rate on one unit of the asset. This modified user cost formula assumes that the specific tax (such as a property tax in the case of a structure or a land plot) is paid at the end of the accounting period. 21. For material on the construction of user costs for more complex systems of business income taxation, see Diewert (1992b) and Jorgenson (1996).

276 W. Erwin Diewert 22. For references to the literature on valuing fixed inputs as they age, see Diewert (2009), Cairns (2013), and Diewert and Fox (2016b). See Diewert and Huang (2011) for a discussion on how to value intellectual property products. Finally, for references to the literature on decomposing property values into their land and structure components, see de Haan and Diewert (2011), Diewert, de Haan, and Hendriks (2015), and Diewert and Shimizu (2015a) for residential property decompositions, Diewert and Shimizu (2017) for condominium property decompositions, and Diewert and Shimizu (2015b, 2016) and Diewert, Fox, and Shimizu (2016) for commercial property decompositions. 23. See Diewert (1980, 475–486) and Schreyer (2001, 2009). 24. The property tax component of the use cost of public-sector property input is also omitted from the SNA user cost treatment. 25. Some countries want to make their GDP as small as possible in order to minimize international transfer payments that are based on their per capita GDP. Thus the treatment of capital services for the general government sector in the System of National Accounts is at least partially a political issue rather than a pure measurement issue. 26. For further discussion on all of these indexes and their properties, see Fisher (1922) and Diewert (1978, 1992a). The US Bureau of Economic Analysis uses chained Fisher indexes to aggregate over inputs and outputs. The use of Törnqvist price and quantity indexes in productivity analysis can be traced back to Jorgenson and Griliches (1967, 1972). Justifications for the use of Törnqvist indexes based on the economic approach to index number theory can be found in Diewert (1976, 1980), Caves, Christensen, and Diewert (1982a, 1982b), Diewert and Morrison (1986), Kohli (1990), and Inklaar and Diewert (2016). Justifications for the use of Fisher indices based on the economic approach to index number theory can be found in Diewert (1992a; 2012, 222–228). 27. Notation: w1 ⋅ x1 ≡ ∑n=1N wn1xn1 denotes the inner product of the vectors w1 and x1. 28. Both of these indexes are superlative indexes; that is, they are exact for flexible functional forms for an underlying flexible functional form for an economic aggregator function as defined by Diewert (1976). Diewert (1978) showed that these two functional forms for an index number formula approximated each other to the accuracy of a second-order Taylor series approximation when the derivatives are evaluated at a point where the two price vectors are equal and where the two quantity vectors are equal. 29. Diewert (1978) pointed this out many years ago. However, chaining does not always work well if the data are available on a subannual basis: seasonal fluctuations and price-bouncing behavior can create a chain drift problem, which was pointed out by Szulc (1983). Ivancic, Diewert, and Fox (2011) and de Haan and van der Grient (2011) suggested adapting multilateral indexes (used to make cross-sectional comparisons) to the time series context in order to deal with the chain drift problem that generally arises when monthly or weekly data are aggregated. See also de Haan and Krsinich (2014). 30. Thus the period t fixed-base implicit quantity indexes that match up with the period t fixed-base Laspeyres and Paasche price indexes are QLX t ≡ w t ⋅ x t /PLX t and QPX t ≡ w t ⋅ x t /PPX t and so on. It should be noted that our is normally called the Paasche quantity index and our QPX t is normally called the Laspeyres quantity index. Our notation is simplified if we label the implicit quantity index that matches up with a particular price index formula in the same way. Note that our QLX t is less than our QPX t for t = 2, . . . , 5. This is a typical relationship between Laspeyres and Paasche price and quantity indexes (but it does not always hold).

Productivity Measurement in the Public Sector 277 31. Diewert (1978) found the same results using Canadian annual national accounts data and thus he advocated the use of chained superlative indexes for national accounts purposes. This is probably good advice if the data are at an annual frequency, but there is the possibility of some chain drift if quarterly data are used. 32. As mentioned earlier, our implicit Laspeyres quantity index QLY t corresponds to what is normally called a fixed-base Paasche quantity index. Cost-weighted input quantity indexes of the type defined by (7.9) have been used widely in the United Kingdom in recent years when constructing measures of non-market output quantity growth; see Atkinson (2005, 88). 33. Generally speaking, gross output TFP growth will be smaller than value-added TFP growth; see Schreyer (2001) and Diewert (2015) for explanations of this phenomenon. 34. See the IMF/Eurostat/ILO/OECD/UNECE/The World Bank (2004). 35. Justifications for this procedure that are based on the economic approach to index number theory can be found in Diewert (1976), Diewert and Morrison (1986), and Kohli (1990), who justified the use of Törnqvist indexes using the economic approach to the measurement of productivity. Diewert (1992a, 2012) justified the use of Fisher indexes using the economic approach. If the public-sector production unit sells some outputs at market prices, then these sales should become a part of the unit’s intermediate input subaggregate, except that the positive signs associated with the quantities of such outputs should be replaced by negative signs. 36. Note that the TFP levels listed in Table 7.9 can also be generated by dividing the input price indexes listed in Table 7.2 by the corresponding real value-added output prices listed in Table 7.7. This result was first established by Jorgenson and Griliches (1967, 252) and is a consequence of the fact that for our example, the value of inputs is always exactly equal to the value of outputs for each period. 37. The reason why TFP = 1 when Laspeyres or Paasche indices are used as the method of aggregation over commodities is due to the fact that these two formulae are consistent in aggregation; that is, if a Laspeyres aggregate is formed by using the Laspeyres formula to aggregate two or more subaggregates and then the resulting subaggregate price and quantities are aggregated in a second stage using the Laspeyres formula again, then the resulting two stage Laspeyres indices of price and quantity are exactly equal to the corresponding Laspeyres indices of price and quantity that are constructed using a single stage of aggregation. On the other hand, the two superlative indices are not exactly consistent in aggregation, but they are approximately consistent in aggregation; see Diewert (1978) for an explanation of these results. 38. We will discuss in more detail the problems associated with measuring health and education outputs in section 7.8. The Atkinson (2005) Report on measuring government outputs in the context of the national accounts has a much more detailed discussion of the associated measurement problems. 39. Data envelopment analysis or benchmarking production units for relative performance was initiated by Farrell (1957) and Charnes, Cooper, and Rhodes (1978) and these methods have been widely applied to public-sector production units. We will not discuss these methods in this chapter since they generally do not generate reasonable output prices in the case where there are many outputs. 40. Scitovsky (1967) suggested this method for imputing prices to outputs produced by the public health sector. Hill (1975, 19–20) noted that unit costs should equal selling prices for competitive market producers and advocated the general use of unit costs to value outputs for non-market producers. Hill carried over his 1975 advice into the System of

278 W. Erwin Diewert National Accounts 1993 where he was a principal contributor. Schreyer (2012a) formally developed the price equals unit cost methodology to value non-market outputs in much more detail. 41. See paragraphs 16.133 and 16.134 of Eurostat, IMF, OECD, UN, and the World Bank (1993). If the public-sector production unit produces some outputs that are sold at market prices, then these outputs can be reclassified as negative intermediate inputs for our purposes; that is, these outputs would appear as negative components in the zt vectors, while the corresponding prices would appear as positive components in the pt vectors. 42. It can be seen that P1 = 1. 43. If it is possible to estimate a period t joint cost function for the public-sector production unit, say Ct(q, p, w), and this cost function is differentiable with respect to the components of the output vector q when evaluated at the period t data so that the vector of first-order partial derivatives ∇qC t (qt , pt , w t ) ≡ P t exists, then this vector of marginal costs can serve as an appropriate vector of cost-based output prices. In addition, if the technology is subject to constant returns to scale, then the value of period t output, P t ⋅ qt , will be equal to period t total cost, C t (qt , pt , w t ) = pt ⋅ z t + w t ⋅ x t . This cost function based methodology for the measurement of the productivity growth of a public-sector production unit is developed in some detail in Diewert (2011, 2012, 2017). 44. The resulting gross output aggregates should be normalized so that aggregate real output equals aggregate real input in the base period.

45. Suppose that the producer’s total cost function for period t, C t (q1t ,…, qK t , pt , w t ) ≡ C t , has been estimated and it is differentiable with respect to the outputs, qk. If in addition, production is subject to constant returns to scale, then the fractions fkt should be defined as f k t ≡ qk t ∂C t (q1t ,…, qK t , pt , w t ) / ∂qk / C t for k = 1, . . . , K. 46. Our suggested methodology is simply an elaboration of a methodology suggested by Schreyer (2010, 21). 47. Sections 7.7 and 7.8 suggest some methods for determining these weights, but there is no generally applicable method for choosing these welfare-oriented weights. In the concluding section of this chapter, we indicate that it is very difficult to determine these valuation weights in a rigorous fashion. 48. Suppose the welfare weights change over time. Thus let ω1 and ω2 be the weights for periods 1 and 2. Define Q1* ≡ 1 and Q 2* ≡ [ω1 ⋅ q 2 /ω1 ⋅ q1 1/2 ω 2 ⋅ q 2 /ω 2 ⋅ q1 ]1/2 . Now use (7.24) to define P1* and P2* and define the normalized prices and quantities by P t ≡ P t * /P 1* and Q t ≡ P 1*Q t * for t = 1,2. The resulting Pt and Qt are Fisher type aggregate prices and quantities for periods 1 and 2. Thus the methodology can be generalized to deal with changing welfare weights. 49. Note that we are forcing the aggregate price times quantity in each period to equal period t total cost; that is, we have P t Q t = P t *Q t * = pt ⋅ z t + w t ⋅ x t for each t. This follows the convention applied to the valuation of non-market production that is recommended by Schreyer (2010, 76): “Throughout this handbook, it is understood that the value of output of institutional units in the health care industry is measured by the observed money value of output in the case of market producers and by the sum of costs in the case of non-market producers. This follows national accounts conventions.” It is not necessary to perform the normalizations of output prices and quantities to make the value of non-market output equal to the total net cost of producing the non-market outputs in period t, but if this is not done, then the non-market production unit will make a profit or loss that is more or

Productivity Measurement in the Public Sector 279 less arbitrary. Thus the Schreyer-Hill normalizing convention leads to a value of aggregate output that will exactly exhaust period t cost. 50. If non-market commodity 1 is not present in period 1 but is present in subsequent periods, then this implies that q11 = 0 and for subsequent periods t > 1 , q1t > 0 . For a more explicit treatment of the new commodity problem in the non-market context, see the following section or Diewert (2012, 220–221). 51. See Atkinson (2005, 88–90) and Schreyer (2012) for nice discussions on the valuation of non-market outputs and the differences between marginal cost and final demander valuations. 52. This problem was pointed out by Griliches (1979, 97): “What happens to price indices will depend on whether they allow for the ‘quality’ improvements embedded in the new item or not. By and large they do not make such quality adjustments. Instead, the new product is ‘linked in’ at its introductory (or subsequent) price with the price indices left unchanged.” Gordon (1981, 130–133) and Diewert (1996, 31; 1998, 51–54) also recognized this source of bias and suggested methods for measuring its magnitude. 53. This analysis is essentially due to Schreyer (2010, 21). Diewert (2012, 220–221) presented a similar analysis of the linking problem that came to the same conclusion. 54. Linking bias occurs not only with respect to the introduction of new products, but also when new lower-cost outlets come into existence. The resulting bias is called outlet substitution bias and it can occur in the public sector as well as in the private sector; see Reinsdorf (1993) and Diewert (1996, 31; 1998, 50–51) for references to the literature. 55. Hausman (1997) estimated these reservation prices econometrically for new breakfast cereals. For a diagram explaining the Hicks methodology, see Diewert (1996, 32). For additional approaches to the estimation of reservation prices, see Lovell and Zieschang (1994), Feenstra (1994), and Diewert (1998, 51–53). 56. For reviews and references to the literature on hedonic regression techniques, see Triplett (1983, 2004), de Haan (2010), de Haan and Diewert (2011), de Haan and Krsinich (2014), and Aizcorbe (2014). Diewert (2011, 180) and Schreyer (2012a, 260–266) present cost function and utility function based hedonic regression models, respectively, to adjust for quality change in the public sector. Schreyer’s model is probably the more appropriate model. 57. Eurostat (2001) also recommended this measure be used by countries belonging to the European Union. 58. “The fte pupils in the four types of maintained school (nursery, primary, secondary and special schools) are added together using cost-weighting by type of school, based on total UK expenditure for that type of school. The cost weights have not been updated since 2000” (Atkinson 2005, 128). He goes on to note that there was a small quality adjustment based on exam success. 59. When measuring the output for a public school class that consists of young students, one could argue that the school is providing a day care component as well as augmenting student skills and knowledge. The day care component of school output would be proportional to student hours spent at school during the time period under consideration. With no change in school hours, student hours would be proportional to school attendance days. Determining non-arbitrary relative weights for the day care and knowledge acquisition components of school output would be problematic. 60. There is a final adjustment that makes the value of output in period t equal to total input cost, as was explained in section 7.6.

280 W. Erwin Diewert 61. This potential volatility could be reduced by making the final quality-adjustment factor equal to an average of the Schreyer quality-adjustment factor that appears in (7.28) and the one that appears in (7.30), where Schreyer’s St could be replaced by our Sk •te . 62. This outcome-oriented methodology is not a new idea. Consider the following quotation from Atkinson (2005, 118): “One approach would be to seek to use weights based on the value of health gain from each treatment rather than on its cost.” 63. “A complete treatment refers to the pathway that an individual takes through heterogeneous institutions in the health industry in order to receive full and final treatment for a disease or condition. . . . Our target definition of health care services includes medical services to prevent a disease” (Schreyer 2010, 73). “The target definition of health care volume output proposed earlier is the number of complete treatments with specified bundles of characteristics so as to capture quality change and new products” (Schreyer 2010, 76). Schreyer goes on to explain why constructing measures of complete treatments is very difficult and hence why one might have to settle for measures of processes that are components of a complete treatment. 64. For an application of the treatment approach in the time series context, see Gu and Morin (2014). 65. Note that the land acquisition costs or opportunity costs for the road can be high. If interest is not allowed as a component of user cost, then the corresponding user cost of the land component of the road will be (mistakenly) set equal to zero and the contribution of public-sector roads to the country’s GDP will be greatly undervalued. It is difficult to determine what an appropriate depreciation rate for the road bed should be, but it will probably be a very small number. The depreciation rate for the asphalt or concrete surface of the road can be relatively large, and it will be related to the utilization of the road. 66. Each customer requires a separate (costly) connection, plus each customer has a separate cost of billing. 67. The exact index number approach to measuring productivity growth can be extended to encompass some limited forms of monopolistic behavior; see Diewert and Fox (2008, 179; 2010, 75). These papers also show that in more general models of monopolistic behavior, exact index number techniques can be used to obtain very simple estimating equations where the contributions of increasing returns to scale and technical progress to productivity growth can be separately identified. 68. See Lawrence and Diewert (2006, 231–233) for an example of how the cost function approach to measuring TFP can be applied in the regulated firm context. For more gen eral expositions of how cost functions can be utilized to measure TFP in the public and regulated sectors, see Diewert (2012) and Schreyer (2012a). 69. Many national income accountants recognize that the current treatment of government- owned capital in the SNA is not consistent with general accounting principles: “The fact that exactly the same kind of service may be provided on both a market and on a non-market basis raises an important question for this report. It is proposed as a matter of principle that the basic methodology used to measure changes in the volume of real output should always be the same irrespective whether the service is provided on a market or on a non- market basis” (Hill 1975, 19). Atkinson (2005, 49) explicitly recommended that the opportunity cost of capital be added to depreciation charges to account for the cost of capital: “We recommend that the appropriate measure of capital input for production and productivity analysis is the flow of capital services of an asset type. This involves adding to the capital consumption an interest charge, with an agreed interest rate, on the entire owned capital.”

Productivity Measurement in the Public Sector 281 70. Alternatively, once f has been determined up to a cardinalization, we could use the resulting time series for ut as our household quantity index. 71. This recovery impossibility theorem does not apply to business demanders of the outputs of a public-sector production unit; it only applies to household demanders. 72. This problem was first noticed by Pollak and Wales (1979, 219). 73. However, this task can be accomplished by making extra assumptions on the structure of the underlying preferences. Adding a time constraint to the household’s budget constraint and making some separability assumptions is one way of proceeding; see Schreyer and Diewert (2014).

References Aizcorbe, A. A. 2014. A Practical Guide to Price Index and Hedonic Techniques. Oxford: Oxford University Press. Atkinson, T. 2005. Atkinson Review: Final Report; Measurement of Government Output and Productivity for the National Accounts. New York: Palgrave Macmillan. Babbage, C. 1835. On the Economy of Machinery and Manufactures, 4th edition. London: Charles Knight. Balk, B. M. 1998. Industrial Price, Quantity and Productivity Indices. Boston: Kluwer Academic. Balk, B. M. 2003. “The Residual: On Monitoring and Benchmarking Firms, Industries and Economies with Respect to Productivity.” Journal of Productivity Analysis 20: 5–47. Cairns, R. D. 2013. “The Fundamental Problem of Accounting.” Canadian Journal of Economics 46: 634–655. Caves, D. W., L. R. Christensen, and W. E. Diewert. 1982a. “The Economic Theory of Index Numbers and the Measurement of Input, Output and Productivity.” Econometrica 50: 1393–1414. Caves, D. W., L. R. Christensen, and W. E. Diewert. 1982b. “Multilateral Comparisons of Output, Input and Productivity using Superlative Index Numbers.” Economic Journal 96: 659–679. Charnes, A., and W. W. Cooper. 1985. “Preface to Topics in Data Envelopment Analysis.” Annals of Operations Research 2: 59–94. Charnes, A., W. W. Cooper, and E. Rhodes. 1978. “Measuring the Efficiency of Decision Making Units.” European Journal of Operational Research 2: 429–444. Christensen, L. R., and D. W. Jorgenson. 1969. “The Measurement of U.S. Real Capital Input, 1929–1967.” Review of Income and Wealth 15: 293–320. Church, A. H. 1901. “The Proper Distribution of Establishment Charges, Parts I, II, and III.” The Engineering Magazine 21: 508–517, 725–734, 904–912. de Haan, J. 2010. “Hedonic Price Indexes: A Comparison of Imputation, Time Dummy and ‘Re-pricing’ Methods.” Journal of Economics and Statistics (Jahrbücher fur Nationalökonomie und Statistik) 230: 772–791. de Haan, J., and W. E. Diewert (eds.). 2011. Residential Property Price Indices Handbook. Luxembourg: Eurostat. de Haan, J., and F. Krsinich. 2014. “Scanner Data and the Treatment of Quality Change in Nonrevisable Price Indexes.” Journal of Business and Economic Statistics 32(3): 341–358. de Haan, J., and H. A. van der Grient. 2011. “Eliminating Chain Drift in Price Indexes Based on Scanner Data.” Journal of Econometrics 161: 36–46.

282 W. Erwin Diewert Diewert, W. E. 1974. “Intertemporal Consumer Theory and the Demand for Durables.” Econometrica 42: 497–516. Diewert, W. E. 1976. “Exact and Superlative Index Numbers.” Journal of Econometrics 4: 114–145. Diewert, W. E. 1978. “Superlative Index Numbers and Consistency in Aggregation.” Econometrica 46: 883–900. Diewert, W. E. 1980. “Aggregation Problems in the Measurement of Capital,” In The Measurement of Capital, edited by D. Usher, 433–528. Chicago: University of Chicago Press. Diewert, W. E. 1983. “The Theory of the Output Price Index and the Measurement of Real Output Change.” In Price Level Measurement, edited by W. E. Diewert and C. Montmarquette, 1049–1113. Ottawa: Statistics Canada. Diewert, W. E. 1992a. “Fisher Ideal Output, Input and Productivity Indexes Revisited.” Journal of Productivity Analysis 3: 211–248. Diewert, W. E. 1992b. “The Measurement of Productivity.” Bulletin of Economic Research 44(3): 163–198. Diewert, W. E. 1996. “Comments on CPI Biases.” Business Economics 32(2): 30–35. Diewert, W. E. 1998. “Index Number Issues in the Consumer Price Index.” Journal of Economic Perspectives 12(1): 47–58. Diewert, W. E. 2005. “Issues in the Measurement of Capital Services, Depreciation, Asset Price Changes and Interest Rates.” In Measuring Capital in the New Economy, edited by C. Corrado, J. Haltiwanger and D. Sichel, 479–542. Chicago: University of Chicago Press. Diewert, W. E. 2008. “OECD Workshop on Productivity Analysis and Measurement: Conclusions and Future Directions.” In Proceedings from the OECD Workshop on Productivity Measurement and Analysis, 11–36. Paris: OECD. Diewert, W. E. 2009. “The Aggregation of Capital over Vintages in a Model of Embodied Technical Progress.” Journal of Productivity Analysis 32: 1–19. Diewert, W. E. 2010. “User Costs versus Waiting Services and Depreciation in a Model of Production.” Journal of Economics and Statistics 230(6): 759–771. Diewert, W. E. 2011. “Measuring Productivity in the Public Sector: Some Conceptual Problems.” Journal of Productivity Analysis 36: 177–191. Diewert, W. E. 2012. “The Measurement of Productivity in the Nonmarket Sector.” Journal of Productivity Analysis 37: 217–229. Diewert, W. E. 2014. “US TFP Growth and the Contribution of Changes in Export and Import Prices to Real Income Growth.” Journal of Productivity Analysis 41: 19–39. Diewert, W. E. 2015. “Reconciling Gross Output TFP Growth with Value Added TFP Growth.” International Productivity Monitor 29(Fall): 17–24. Diewert, W. E. 2017. “Productivity Measurement in the Public Sector: Theory and Practice.” Discussion Paper 17–01, Vancouver School of Economics, University of British Columbia. Diewert, W. E., J. de Haan, and R. Hendriks. 2015. “Hedonic Regressions and the Decomposition of a House Price Index into Land and Structure Components.” Econometric Reviews 34(1–2): 106–126. Diewert, W. E., and K. J. Fox. 2008. “On the Estimation of Returns to Scale, Technical Progress and Monopolistic Markups.” Journal of Econometrics 145: 174–193. Diewert, W. E., and K. J. Fox. 2010. “Malmquist and Törnqvist Productivity Indexes: Returns to Scale and Technical Progress with Imperfect Competition.” Journal of Economics 101: 73–95. Diewert, W. E., and K. J. Fox. 2016a. “Alternative User Costs, Rates of Return and TFP Growth Rates for the US Nonfinancial Corporate and Noncorporate Business Sectors: 1960–2014.” Discussion Paper 16–03, Vancouver School of Economics, University of British Columbia.

Productivity Measurement in the Public Sector 283 Diewert, W. E., and K. J. Fox. 2016b. “Sunk Costs and the Measurement of Commercial Property Depreciation.” Canadian Journal of Economics 49: 1340–1366. Diewert, W. E., K. J. Fox, and C. Shimizu. 2016. “Commercial Property Price Indexes and the System of National Accounts.” The Journal of Economic Surveys 30: 913–943. Diewert, W. E., and N. Huang. 2011. “Capitalizing R&D Expenditures.” Macroeconomic Dynamics 15(4): 537–564. Diewert, W. E., and C. J. Morrison. 1986. “Adjusting Output and Productivity Indexes for Changes in the Terms of Trade.” The Economic Journal 96: 659–679. Diewert, W. E., and A. O. Nakamura. 2003. “Index Number Concepts, Measures and Decompositions of Productivity Growth.” Journal of Productivity Analysis 19: 127–159. Diewert, W. E., A. O. Nakamura, and L. Nakamura. 2009. “The Housing Bubble and a New Approach to Accounting for Housing in a CPI.” Journal of Housing Economics 18(3): 156–171. Diewert, W. E., and C. Shimizu. 2015a. “Residential Property Price Indices for Tokyo.” Macroeconomic Dynamics 19: 1659–1714. Diewert, W. E., and C. Shimizu. 2015b. “A Conceptual Framework for Commercial Property Price Indexes.” Journal of Statistical Science and Application 3(9–10): 131–152. Diewert, W. E., and C. Shimizu. 2016. “Alternative Approaches to Commercial Property Price Indexes for Tokyo.” Review of Income and Wealth 63(3): 492–519. Diewert, W. E., and C. Shimizu. 2017. “Hedonic Regression Models for Tokyo Condominium Sales.” Regional Science and Urban Economics 60: 300–315. Diewert, W. E., and H. Wei. 2017. “Getting Rental Prices Right for Computers.” Review of Income and Wealth 63(Supplement 1): S149–S168.. Eurostat. 2001. Handbook on Price and Volume Measures in National Accounts. Brussels: European Commission. Eurostat, IMF, OECD, UN, and the World Bank. 1993. System of National Accounts 1993. New York: United Nations. Farrell, M. J. 1957. “The Measurement of Production Efficiency.” Journal of the Royal Statistical Society, Series A 120: 253–278. Feenstra, R. C. 1994. “New Product Varieties and the Measurement of International Prices.” American Economic Review 84: 157–177. Fisher, I. 1922. The Making of Index Numbers. Boston: Houghton-Mifflin. Fox, K. J., and U. Kohli. 1998. “GDP Growth, Terms of Trade Effects and Total Factor Productivity.” Journal of International Trade and Economic Development 7: 87–110. Griliches, Z. 1979. “Issues in Assessing the Contribution of Research and Development to Productivity Growth.” Bell Journal of Economics 10 (Spring): 92–116. Griliches, Z., and I. Cockburn. 1994. “Generics and New Goods in Pharmaceutical Price Indexes.” American Economic Review 84: 1213–1232. Gordon, R. 1981. “The Consumer Price Index: Measuring Inflation and Causing It.” The Public Interest 63 (Spring): 112–134. Gu, W., and S. Morin. 2014. “Experimental Measures of Output and Productivity in the Canadian Hospital Sector, 2002 to 2010.” Canadian Productivity Review, Catalogue No. 15- 206-X–No. 034. Ottawa: Statistics Canada. Harper, M. J., E. R. Berndt, and D. O. Wood. 1989. “Rates of Return and Capital Aggregation Using Alternative Rental Prices.” In Technology and Capital Formation, edited by D. W. Jorgenson and R. Landau, 331–372. Cambridge MA: MIT Press.

284 W. Erwin Diewert Hausman, J. 1997. “Valuation of New Goods under Perfect and Imperfect Competition.” In The Economics of New Goods, edited by T. Bresnahan and R. Gordon, 209–237. NBER Studies in Income and Wealth, Volume 58. Chicago: University of Chicago Press. Hicks, J. R. 1940. “The Valuation of the Social Income.” Economica 7: 105–140. Hill, P. 1975. Price and Volume Measures for Non-Market Services. Brussels: Statistical Office of the European Communities. Hill, R. J., and T. P. Hill. 2003. “Expectations, Capital Gains and Income.” Economic Inquiry 41: 607–619. Inklaar, R., and W. E. Diewert. 2016. “Measuring Industry Productivity and Cross-Country Convergence.” Journal of Econometrics 191: 426–433. IMF/Eurostat/ILO/OECD/UNECE/The World Bank. 2004. Producer Price Index Manual: Theory and Practice, edited by Paul Armknecht. Washington, DC: International Monetary Fund. Ivancic, L., W. E. Diewert, and K. J. Fox. 2011. “Scanner Data, Time Aggregation and the Construction of Price Indexes.” Journal of Econometrics 161: 24–35. Jorgenson, D. W. 1989. “Capital as a Factor of Production.” In Technology and Capital Formation, edited by D. W. Jorgenson and R. Landau, 1–35. Cambridge MA: MIT Press. Jorgenson, D. W. 1996. Investment, Vol.2: Tax Policy and the Cost of Capital. Cambridge, MA: MIT Press. Jorgenson, D. W., and Z. Griliches. 1967. “The Explanation of Productivity Change.” The Review of Economic Studies 34: 249–283. Jorgenson, D. W., and Z. Griliches. 1972. “Issues in Growth Accounting: A Reply to Edward F. Denison.” Survey of Current Business 52(4) Part II (May): 65–94. Kohli, U. 1990. “Growth Accounting in the Open Economy: Parametric and Nonparametric Estimates.” Journal of Economic and Social Measurement 16: 125–136. Koechlin, F., P. Konijn, L. Lorenzoni, and P. Schreyer. 2015. “Comparing Hospitals and Health Prices and Volumes across Countries: A New Approach.” Social Indicators Research (December): 1–22. Lawrence, D. and W. E. Diewert. 2006. “Regulating Electricity Networks: The ABC of Setting X in New Zealand.” In Performance Measurement and Regulation of Network Utilities, edited by T. Coelli and D. Lawrence, 207–241. Cheltenham, UK: Edward Elgar. Lovell, C. A. K., and K. Zieschang. 1994. “The Problem of New and Disappearing Commodities in the Construction of Price Indexes.” In Data Envelopment Analysis: Theory, Methodology and Application, edited by A. Charnes, W. W. Cooper, A. Y. Lewin and L. M. Seiford, 353–369. New York: Springer. Peasnell, K. V. 1981. “On Capital Budgeting and Income Measurement.” Abacus 17(1): 52–67. Pollak, R. A., and T. J. Wales. 1979. “Welfare Comparisons and Equivalence Scales.” American Economic Review 69(2): 216–221. Reinsdorf, M. 1993. “The Effect of Outlet Price Differentials in the U.S. Consumer Price Index.” In Price Measurement and Their Uses, edited by M. F. Foss, M. E. Manser, and A. H. Young, 227–254. NBER Studies in Income and Wealth, Volume 57. Chicago: University of Chicago Press. Schreyer, P. 2001. Measuring Productivity: Measuring Aggregate and Industry Level Productivity Growth. Paris: OECD. Schreyer, P. 2009. Measuring Capital: OECD Manual 2009, 2nd edition. Paris: OECD. Schreyer, P. 2010. “Towards Measuring the Volume Output of Education and Health Services: A Handbook.” OECD Statistics Working Paper No. 31. Paris: OECD.

Productivity Measurement in the Public Sector 285 Schreyer, P. 2012a. “Output, Outcome and Quality Adjustment in Measuring Health and Education Services.” The Review of Income and Wealth 58(2): 257–278. Schreyer, P. 2012b. “Measuring Multifactor Productivity When Rates of Return Are Endogenous.” In Price and Productivity Measurement, Vol. 6: Index Number Theory, edited by W. E. Diewert, B. M. Balk, D. Fixler, K. J. Fox, and A. O. Nakamura, 13–40. Victoria: Trafford. Schreyer, P., and W. E. Diewert. 2014. “Household Production, Leisure and Living Standards.” In Measuring Economic Sustainability and Progress, edited by D. W. Jorgenson, J. S. Landefeld, and P. Schreyer, 89–114. Chicago: University of Chicago Press. Scitovsky, A. A. 1967. “Changes in the Cost of Treatment of Selected Illnesses, 1951–65.” American Economic Review 57: 1182–1195. Szulc, B. J. (Schultz). 1983. “Linking Price Index Numbers.” In Price Level Measurement, edited by W. E. Diewert and C. Montmarquette, 537–566. Ottawa: Statistics Canada. Triplett, J. E. 1983. “Concepts of Quality in Input and Output Price Measures: A Resolution of the User Value and Resource Cost Debate.” In The U.S. National Income and Product Accounts: Selected Topics, edited by M. F. Foss, 269–311. NBER Studies in Income and Wealth, Volume 47. Chicago: University of Chicago Press. Triplett, J. E. 2004. “Handbook on Hedonic Indexes and Quality Adjustments in Price Indexes: Special Application to Information Technology Products.” STI Working Paper 2004/9. Paris: OECD. Walras, L. 1954. Elements of Pure Economics, a translation by W. Jaffé of the Edition Définitive (1926) of the Eléments d’économie pure, first edition published in 1874. Homewood, IL: Richard D. Irwin.

Chapter 8

Produ ct i v i t y Measurem e nt a nd the Environme nt Finn R. Førsund

8.1. Introduction The purpose of the chapter is to review and explore attempts to measure productivity when facing environmental problems caused by economic activities of production and consumption. The emphasis will be on how to include degradation of the natural environment by the discharge of pollutants from these activities. We are facing both regional and global negative effects, such as acid rain and global warming, the latter leading to sea level rise, forced human migrations, and adverse health effects, such as diseases, including malaria and the zika virus, caused by the spread of mosquitoes to new areas. Another aspect of environmental degradation not reflected in how conventional productivity is measured is the depletion of natural resources. The value of extracting or harvesting commercial resources is entered into national accounts, but the running down of stocks is not. A problem with the national accounts is that only current market transactions are included; the valuation of resource depletion and environmental degradation of the public good aspects of Nature is absent. This chapter will focus on how to incorporate environmental degradation, but will not attempt to deal in depth with resource depletion. Environmental problems until the late 1950s had been treated by economists as cases of externalities. The examples used had an innocent flair; Pigou (1920) used factory smoke dirtying laundry hanging to dry outdoors as a negative externality, and Meade (1952) used the interaction of bees and apple blossoms as a positive externality. The publishing of the seminal paper by Ayres and Kneese (1969) (see also Kneese et al. 1970; Kneese 1971), coining the phrase materials balance, heralded a new view within economics of the pervasiveness and seriousness of environmental pollution.1 The

288 Finn R. Førsund conservation of mass and energy (based on the first law of thermodynamics) tells us that matter cannot be created, nor can it disappear. If all the material inputs into an activity are not embedded in the products the activity is set up to deliver, then the difference must be contained in residuals discharged to the environment. In other words, if we weigh the inputs employed in an activity, and weigh the products that are the purpose of activities, the difference is the residuals (taking into consideration nonpaid factors like oxygen from the air), which may turn out to be polluting the natural environment. Thus, the general feature of residuals is that they arise from the use of material inputs in a wide sense. The concept of materials balance underlines the general inevitability of residuals generation and the pervasiveness of pollution when employing material resources. The materials balance principle leads naturally to a model of production activity (by firms or consumers) based on joint production of the intended goods of the production activity and unintended products. The latter are in general termed residuals, but they become pollutants if their discharge to the environment leads to degradations as evaluated by consumers. The plan of the chapter is as follows. Section 8.2 introduces briefly the conventional way of measuring productivity as adopted when using data from the national accounts. Macro issues such as dynamics of growth with pollutants, sustainability, green national product, and the extension of input—output analysis to cover pollutants are touched upon. This section introduces physical satellite accounts for resource depletion and refers to the state of the environment. The concept of an environmental damage function is defined as the willingness to forsake man-made goods in order to sustain a certain quality of the environment. This cost is a key term in environmental economics. The effect on productivity of introducing environmental costs will then be shown in principle. In Section 8.3 the micro-level model used for studying the interaction of production and generation of pollution is introduced. The key assumption is that the use of material inputs in the production of intended (desirable) goods leads simultaneously to the generation of residuals that may turn out to be polluting the natural environment. The model consists of two types of relations: a production function for the intended output, and another production function for the unintended output. This model is consistent with the materials balance discussed in detail in Subsection 8.3.2; there is no trade-off between desirable and undesirable outputs for given resources, as implied by the material balance and efficiency assumptions on technology (Førsund 2018). Technical change is defined, and end-of-pipe abatement is introduced as a separate activity. The social optimization problem using a damage function representing the evaluation of environmental degradation is set up in Section 8.4, and the impact of direct environmental regulation on productivity and profitability is shown. The Porter hypoth esis of positive technology effects of environmental regulation is studied in Section 8.5. Basic regulatory instruments such as the Pigou tax and cap and trade are studied. The key statements that environmental regulation represents a pressure on firms to reduce short-term inefficiency, and in the longer run leads to innovation reducing the generation of pollutants, are discussed. The impact on productivity change is commented upon. Productivity measures assuming persistent inefficiency of some firms, compared

Productivity Measurement and the Environment 289 with the contemporaneous best-practice frontier function, are discussed in Section 8.6. It is argued that the most popular model in the inefficiency literature applied to pollution from the late 1980s is based on a single-equation model that does not satisfy the materials balance, and efficiency of resources utilization. Consequently, multi- equation models should be used instead. A summary of the chapter’s main results and conclusions is provided in Section 8.7, and indications of further research issues are offered.

8.2. Productivity Measures 8.2.1. Conventional Productivity Measures Productivity may be defined in several ways and used both in a macro and micro setting.2 The most common definition is to use some measure of production per unit of labor used as an input in the activity in question; productivity can be measured simply as Y/L where Y is the output and L the input of labor. The two output measures used based on the national accounts are value added and gross production. Value added is the difference between revenue and variable cost excluding labor and capital, and gross production is measured as shipments, sales, or revenue. Value added is the preferred measure at an aggregated level. In the national accounts it can be calculated from the supply side as defined above, or from the income side. Then value added is the sum of labor expenditures and remuneration to capital, including depreciation and return on the assets (including owner income). At an industry level or lower levels of aggregation, a multifactor measure of productivity based on gross output may be preferred. Then an index of inputs is used in the denominator. Gross output in the numerator also has to be represented by an index if there is more than one output. We must distinguish between productivity and productivity change. The latter is the relative change in productivity from one period to another. The seminal definition of productivity change in continuous time is the Solow residual (Solow 1956), where productivity change is the same as change in technology, and is defined as the growth in output that cannot be explained by the growth of inputs. The labor productivity measure is a partial measure. Looking at the contribution of all factors of production, a concept of multifactor productivity can be defined in the single output case as

TFP =

Y (8.1) F (x K , x L , x E , x M , x S )

where the F(.) function is an aggregator function, having positive partial derivatives and being linear homogenous, in the typical factors xj (j = K, L, E, M, S) with the subindices denoting production capital K, labor L, energy E, materials M, and services

290 Finn R. Førsund S. In Organisation for Economic Co-operation and Development (OECD) and EU studies of production, KLEMS is used as a term for these factors.3 Multifactor productivity reflects the change in output that cannot be accounted for by the change in all the inputs. The best output measure corresponding to the use of all inputs is gross output. We can aggregate the individual outputs yielding revenue using an output aggregator function. A Törnqvist index of productivity on logarithmic form uses the revenue shares as weights in a linear aggregation of outputs and cost shares to aggregate inputs, using the behavioral assumption of revenue maximization or cost minimization.

8.2.2. Extending Productivity Measures to Include the Environment The problem of pollution as a byproduct of economic activity is a major topic in contemporary environmental economics, ranging from global warming due to generation of greenhouse gases, to deteriorations in local air, water, and land quality due to emission and discharge of a variety of polluting substances, as pointed out in Section 8.1. Other related issues both on the macro and the micro level should also be mentioned. At the macro level the interest was focused on the dynamics of growth, including pollutants of the accumulating kind (d’Arge and Kogiku 1973; Keeler et al. 1972; Nijkamp and Paelinck 1973; Plourde 1972; Smith 1972). The Pontryagin’s maximum principle was the method adapted.4 Resource depletion was also analyzed at the aggregated level with this new mathematical tool, following up on the early work of Hotelling in the 1930s. As stated in Section 8.1, this is an environmental problem of its own, but it will not be treated here. The interest in sustainable growth led to some specific principles for utilizing natural resources including environmental ones. Hartwick’s rule (Hartwick 1977) for compensating resource depletion was to invest the resource rent. This was called genuine savings. Solow (1974) showed that one way to design a sustainable consumption program for an economy is to accumulate produced capital sufficiently rapidly. The pinch from the shrinking exhaustible resource stock can then be precisely countered by the services from the enlarged produced capital stock, given a sufficient degree of substitutability between produced capital and natural resources (see also Mäler 1991; Weitzman 1976 on the savings rule). Since the construction of national accounts became common in most countries after World War II (through the work of the United Nations Statistical Division, UNSD), there has been a discussion about gross national product (GNP) (or gross domestic product; GDP) as a welfare measure (Stiglitz et al. 2009). The blossoming of environmental economics also led to a wish to have the (mis)use of environmental resources included in a welfare measure, as well as the running down of exhaustible resources. Indeed, there is also a more recent interest in having national accounts provide a measure of gross human happiness, which would be a comprehensive welfare measure that also includes

Productivity Measurement and the Environment 291 environmental concerns. (However, the only country that has attempted to make a human happiness account is Bhutan.) The term green national product (see, e.g., Aaheim and Nyborg 1995; Asheim 2000) has been used to refer to attempts to integrate national accounting with environmental considerations. A key concept is sustainable consumption that can be based on the Hicksian definition of sustainable income as consumption that can be had without degrading the resource base or environmental assets. Attempts to integrate national accounts and the emission of pollutants were also tried at a more detailed level of aggregation for input–output models (Leontief 1970, 1974; Leontief and Ford 1972). An abatement sector dealing with pollutants was introduced in the papers. The fixed input–output coefficients were extended to include fixed- emission coefficients for various pollutants, calculated as emissions per unit of output. Recognizing the role of material inputs, fixed coefficients related to outputs assume that there are fixed coefficients in production in general, as there are in the input– output model. Based on data for Norway, the costs of obtaining a “greener” mix of final deliveries for a given amount of primary inputs were shown in Førsund and Strøm (1976) and Førsund (1985). However, it turned out to be too difficult to integrate environmental issues formally and statistically in the national accounts. Instead, so-called satellite accounts for natural resources, environmental resources, and pollution emissions were established (United Nations 1993; Nordhaus and Kokkelenberg 1999). These accounts keep track of the extraction, harvesting, and development of stock of resources, and record emissions and change in various quality indicators of the natural environment. We will limit our attention to the micro level of production, and for ease we use a firm as the term for a production unit. The effect of environmental regulation on various aspects such as productivity and profitability in the short and long run will be considered. The point of departure of this chapter is the way of modelling the interaction between economic activity and the natural environment within environmental economics (Førsund and Strøm 1988; Perman et al. 2011). The standard productivity indices used both on the macro level and on a micro level have to be extended to include environmental concerns. How, then, may productivity measures be extended to include environmental assets? The general approach is to express environmental considerations in measuring units compatible with the way outputs are evaluated (i.e., money). Nature provides us with a number of amenities or services: clean air and water, wildlife, biodiversity, recreational possibilities, and aesthetic experiences, among many other forms of services. However, it should be realized that it is not meaningful to ask what the value or price of such services is in an absolute sense. The point is that a valuation must be based on what we are willing to give up of man-made goods to enjoy such services at specific quality levels, rather than being based on any concept of intrinsic value. We will narrow down our approach to investigate pollutants generated as byproduct of economic activity. A key construct in environmental economics is to use a damage

292 Finn R. Førsund function monetizing the degradation of the environment by discharging pollutants into Nature:

D(z ) = D(z1 ,..., z R ) ,

∂D ≥ 0, ∂zi

i = 1,..., R (8.2)

where z is a vector of R types of pollutants. If we have ∂D(z ) / ∂zi > 0, the residual is called undesirable or a pollutant. This will be assumed to be the case in the following. The damage is measured as the value we are willing to give up (or demand in compensation) of produced goods, the production of which generates the pollutants as unintended products. Damage cost will be used synonymously with environmental costs. Nature has a spatial dimension. A damage function has to be detailed a lot more before we can talk about how to estimate the damage function. A distinction has to be made between local pollution from a single source and the transport of pollutants from several sources to receptors in Nature where the damage is occurring. The latter form is called regional pollution. Damage depends on the location of the sources, and the impacts of the emission from each source can in principle be identified. Acid rain is a typical example, and has been subject to large-scale modelling effort at the International Institute for Applied System Analysis (IIASA) initiated by the United Nations International Economic Commission for Europe (UNECE). Acid substances are transported via air from the sources, and the place where it is deposited on the ground or vegetation may be a long way from the source, especially if high chimneys are used or flue gases are ejected using pressurized air. Another main type of damage function is emission to the air that mixes in the atmosphere and becomes a global pollutant as climate gases, causing climate change. The type of gas is then simply summed over all sources emitting this gas. Although the damages are local, all sources contribute to these damages. A technique developed within environmental economics is to measure the willingness to pay for preserving the present environmental qualities, or willingness to pay for an improvement. Necessary compensation demanded for accepting deterioration in environmental quality is also used. The damage function (8.2) is a theoretical construct commonly used in textbooks on environmental economics based on such willingness- to-pay concepts. Methods of estimation may be based on transport costs of visiting recreation sites, complementary equipment bought to enjoy recreation activities, or contingent valuation based on elaborated interview schemes (see, e.g., Førsund and Strøm 1988; Perman et al. 2011). For a given level of aggregation (e.g., using the industrial classification system of national accounts) let pi be the price and yi the volume of desirable good i. Measuring the social gross output, the damage cost of the environmental services caused by producing the goods must now be subtracted from the revenue of desirable goods to be used in the numerator, calculating the social total factor productivity TFPE:

TFPE =

∑ py i

i

i

− D(z )

F (x K , x L , x E , x M , xS )

(8.3)

Productivity Measurement and the Environment 293 In a static setting, it is obvious that the social multifactor productivity must be lower because of the reduction in the numerator, assuming no change in outputs and inputs (i.e., TFPE < TFP). However, once environmental damage is included in the calculation of social gross output in a period predating period t, the productivity change may go either way:

∆TFPEt ,t +1 =

(

∑

t +1 K

i

pi yit +1 − D(z t +1 ) t +1 L

t +1 E

t +1 M

F x ,x ,x ,x ,x

t +1 S

)

∑

(

t K

i

pi yit − D(z t ) t L

t E

t M

F x ,x ,x ,x ,x

t S

)

> 1 (8.4)

0 (8.5) z = g (x M , x S ), g x′M > 0, g x′S ≤ 0

Here y is the desirable output that it is the purpose of setting up the activity, xM and xS are material inputs and service inputs, respectively, and z is the unintended residual generated simultaneously, utilizing inputs to produce outputs. The multi-equation system of Frisch consists of a production function f(.) for each desirable output as a function of the same bundle of inputs, and a production function g(.) for the generation of each residual as a function of the same set of inputs. This is product separation. The functions f(.) and g(.) are assumed to be efficient in the sense that y is maximized for a given bundle of inputs, while the residual z is minimized for the same given bundle of inputs, so equalities are used in (8.5).6 The system of equations (8.5) is a drastic simplification of engineering realities, but still captures the most essential feature of the type of joint production of goods and residuals; the crucial connection between these two types of outputs goes through the use of material inputs. It is not the case that the desirable output in general has a fixed relationship with the undesirable one independent of the inputs. The environmental problem of generation of residuals is that emitting them to the natural environments may create negative externalities due to degradation of environmental qualities or in general creating harmful effects. As defined following Equation (8.2), residuals are then called pollutants. To keep the model as simple as possible, we consider a single desirable output y and an undesirable output z (generalizing to multi-output and multi-pollutants can be done in the Frisch model just by adding more equations, one for each variable, keeping the same inputs as arguments in all relations; see Førsund 2009). The material inputs are fossil primary energy (in the form of, e.g., coal, oil, gas, and wood), and various types of raw materials. The service inputs are those that are not used up in the production process in a material sense, remaining physically intact and providing services such as labor and capital. Electricity used as input is in our context also a service input since it does not have any mass. The production of desirable outputs and undesirable residuals occurs simultaneously and is based on the technology of multiple-output production. The separation into two types of equations

Productivity Measurement and the Environment 295 does not mean that we have two technologies, but is done as a very helpful simplification of a possibly complex technology, without sacrificing key aspects of a technology such as identifying substitution between inputs, scale properties, and the connection between desirable and undesirable outputs. The partial derivatives (the productivities) of the desirable output function are assumed to have the usual positive signs. However, concerning the production function for the undesirable output, the two types of inputs are assumed to have opposite signs: marginal increases in material inputs increase the undesirable output, while marginal increases in service inputs decrease the undesirable output. The positive partial productivity of service inputs in the desirable output production function and the negative sign in the residuals generation function can be explained by the fact that more of a service input improves the utilization of the given raw materials through better process control or increased internal recycling of waste materials.7 As a consequence of the typical way in which service inputs increase the output by utilizing the material inputs more efficiently, the production of pollutants will then typically decrease. Therefore, the signing of the marginal productivity in the pollutants production function is negative. However, if the output y is non-material then g x′S = 0. The material inputs are essential in the sense that we will have no production, either of goods or pollutants, if xM is zero:8

y = f (0, x S ) = 0, z = g (0, x S ) = 0 (8.6)

There will in general be substitution possibilities between material and service inputs. The rate of substitution evaluated at a point on an isoquant of the production function for the desirable output is (− f x′M / f x′S ) < 0. This is the amount of material input that is reduced if the service input is increased with one unit, keeping output y constant. Considering several material inputs, there may be substitution possibilities between them, for example between coal and natural gas, keeping the output constant, but decreasing the generation of bads if the marginal contribution of gas to the creation of bads is smaller than the marginal contribution of coal: g x′coal > g x′ gas for the same heat content. The role of service inputs in the residuals production function is crucial as to the substitution effect of decreasing a material input and increasing a service input, resulting in less generation of residuals for a constant production of the desirable output. The marginal rate of substitution is positive, (− g x′ M / g x′S ) > 0, due to the marginal productivity of service inputs being negative. This implies a special form of isoquants in the factor space and the direction of increasing residual level compared with a standard isoquant map for the output, as seen in Figure 8.1. The isoquants for the two outputs can be shown in the same diagram because the arguments in the functions are the same (see Frisch 1965, 272; Førsund 2009, 7) for the original illustration of isoquants in the factorially determined multi-output case). The level of the residual z is increasing moving southeast (see dotted arrow pointing

296 Finn R. Førsund xS

z B C

D xS

A

A

xM A

y

xM

Figure 8.1. Isoquants for the production of y and z.

southeast) while the level of the desirable good y is increasing moving northeast (see dotted arrow pointing northeast). Going from point A to point B in input space increasing both inputs, but nonproportionally, we see that the production of the residual z has decreased while the production of output y has increased. Going from B to C reducing the input xS and increasing the input xM the desirable output is constant, but the undesirable output has increased. The two types of outputs y and z are only implicitly related through being generated by the same inputs as shown by the model (8.5); there is no trade-off between y and z for given x. In the input–output space we only get one point each for y and z given a bundle of inputs x as implicitly shown by, for example, points A and B where the isoquants for the two outputs intersect. There are obviously limits to substitution between material and service input keeping the same desirable output. Moving along the y isoquant from point A in a northwest direction, there is a limit to the amount of raw materials that can be extracted from the material input while keeping the output constant (i.e., there is a lower limit on how much the residual generation can be reduced). Another angle on this lower limit is keeping the material input constant at the level xMA at point A, and then see how much residuals can be reduced increasing the service input from xSA. Let us say point D will be the point with the minimum generation of residuals z, but then the good output has also increased. The minimum level of residual generation depends on the level of both types of inputs, as does the level of maximal good output.

8.3.2. The Materials Balance Assuming that one or more inputs to a production process of a production unit consists of physical mass, this mass will not disappear during the production process, but, as stated previously in Section 8.1, must either be contained in the products being produced, or become residuals emitted to the external environment. Thus, for model (8.5) a materials balance exists for each production unit, and α, β, γ are coefficients converting units of inputs x and desirable outputs y and residuals z to a common mass unit:

αx M ≡ β y + γ z (8.7)

Productivity Measurement and the Environment 297 As in the earlier discussion of (8.5), we simplify for convenience by operating formally with only a single input, desirable output and residual, respectively, but generalization to multiple inputs, outputs, and residuals is a straightforward, but cumbersome extension of the number of equations and a summation over the type of each variable yielding the same pollutant. Regarding combustion processes, it is the case that substances drawn from the air like oxygen and nitrogen may combine with substances in the raw materials (e.g., creating CO2, NOx, SO2, etc.). The mass balance can either reflect these both on the left-hand and right-hand side of (8.7), or the residual can be considered as the substance contained in the raw material only, like carbon or sulfur. We follow the last option. Assuming that inputs and desirable outputs are homogenous across units, the coefficients α and β must be equal for each type of input and output, and assuming that given homogenous inputs, and respectively desirable outputs, will generate homogenous undesirable outputs, the coefficient γ is also constant across units. The parameters α, β and γ are technical unit-conversion coefficients and are not parameters of a production relation. It should therefore be realized that the materials balance is not a production function relation proper. All observations of a production activity, perfectly observed, must obey the materials balance as a physical law. That is why an identity sign is used in (8.7). It is without good meaning, for example, to differentiate (8.7) w.r.t. y and z holding x constant using (8.7) as a production relation. If (8.7) is regarded as a production technology, y cannot increase without increasing x; a given input x cannot be reallocated from producing z to produce y. If a parametric or nonparametric efficient production function is estimated based on observations of x, y, and z, the materials balance must also hold for unobserved points obeying the production function. In this sense the materials balance puts a constraint on the efficient production function construction.9 However, the materials balance represents another type of link between the two types of outputs and inputs than the relations representing the production technology. The materials balance does not tell us anything about the specific technology that has generated the observations of the variables in (8.7) (Førsund 2018). Desirable outputs y cannot be produced without residuals z if the complete amount of the material content of the inputs is not contained in the outputs. This is the inevitability of generating pollutants together with the desirable product. Thermal electricity generation is an extreme case with no material content of the desirable output, implying β = 0, and then all material content of the material input ends up as waste products. The aging of a vintage wine may be an extreme example of an opposite case with zero residual. All the material content of the vintage wine is contained in the desirable output, the wine bottle; if we assume that no other material inputs are required, such as heating/cooling of the storage facility (e.g., the wine is stored in a natural cave with the right temperature). (However, the production activity is just transportation in time with chemical processes taking place within the same mass.) The materials balance is fulfilled for any combination of inputs in the model (8.5) because of the unique correspondence between input use and outputs generated by these

298 Finn R. Førsund inputs. The same inputs determine both outputs, so we only get two single values for the outputs for a given level of inputs. Then the materials balance (8.7) being a physical law is also obeyed by definition. Therefore, in our type of model, the materials balance can be regarded as a pure accounting identity since it holds at any point in input–output space satisfying the production functions. This is illustrated in Figure 8.1. At point A or B, the desirable output cannot be increased by reducing the undesirable output keeping inputs constant; all points giving the values in output space are intersections of isoquants in the input space. The inputs supporting point A give simultaneously the amounts of desirable and undesirable outputs given by the intersection point of the respective isoquants, and similarly at point B. Moving from A to B increases the desirable output and decreases the undesirable output, but this is happening due to an increase in material inputs and a considerable increase in service input. Model (8.5) does not allow any transformation relationship between the desirable and undesirable outputs for a given level of inputs in the usual sense of a trade-off between the desirable and the undesirable outputs for a given bundle of inputs in a single- equation system.

8.3.3. Technical Change There are four main methods of influencing the generation of undesirable output when keeping the same type of desirable output. Two methods can be done within a static technology, and two other methods require change in technology. Output of desirable goods may be reduced in such a way that the use of material inputs is also reduced. This is the only option at a sectoral level within a standard input–output model with fixed coefficients in production (see Subsection 8.2.2). Another way is to utilize substitution possibilities between material inputs with different marginal productivities in the residuals production function (e.g., switch from coal to natural gas in thermal generation of electricity), and between material inputs and service inputs. These substitution options are captured by our model (8.5). A third way is to change the production technology so as to create less pollution for a constant output of desirable goods. Technology improvements that may be small-scale and introduced in the short run (e.g., a period of a year or less) may be considered variable factors.10 But technology improvement may also need large capital investments; changing the main production processes into technologies that use fewer amounts of raw materials, or processing them in such a way that less waste of material inputs occur. Such changes will be more of a long-term character based on real capital with a long technical lifetime. This type of technical change is typical in industries using capital with embodied technology. Investment in new technology developed to use fewer raw materials or to reduce waste is needed in order to reduce generation of residuals.

Productivity Measurement and the Environment 299 Technology change, covering both short run and long run, means a simultaneous change of the functional forms f(.) and g(.) over time:

f t2 (x M , x S ) > f t1 (x M , x S ) (8.8) g t2 (x M , x S ) < g t1 (x M , x S ), t 2 > t1

Technical change in the two production functions in (8.5) can be illustrated in Figure 8.1 by just changing the level labeling of the two sets of isoquants. A positive technical change of the desirable output production function means producing more for given inputs, while positive “green” technical change in the residuals production function means generating less residuals for the same input levels. Note that the unit conversion coefficients in the materials balance equation (8.7) remain the same as long as the inputs and the outputs stay the same. However, technical change may also involve using new types of inputs or redesigning products to use fewer raw materials (“dematerialization” of desirable outputs). These possibilities are not explored further here. A fourth possibility is to install a separate facility using the residuals from (8.5) as inputs and processing them in such a way that less harmful pollutants result, for example capturing particles using electrostatic filters on smoke stacks converting an air pollution into a solid waste problem, and installing scrubbers converting emission to air to discharges to water. Such facilities are called end-of-pipe abatement in environmental economics and will be addressed in the next subsection.

8.3.4. End-of-Pipe Abatement We will add a specific abatement process to the multi-equation model (8.5) that may need inputs distinct from inputs employed in the production process described by the equations in (8.5). End-of-pipe abatement often consists of a facility separated from the production activity. Another abatement option in the short run is to retool the processes and do small-scale changes, as mentioned earlier. This option is an alternative to integrated technological process solutions. However, it is often rather difficult to identify such activities and to identify the inputs involved. It is easier to do this with a stand- alone abatement facility in terms of inputs used and outputs produced. The residuals generated in the production process will then be channeled to a treatment facility and will be regarded as inputs in this activity. In addition, other inputs—like labor, capital, chemicals, absorbing substances, and energy—may have to be used in order to convert part of the original pollutants into outputs being abated pollutants, creating less harm (usually assumed to be zero) than the primary ones (Førsund 2009). In the long run there may be a choice between end-of-pipe abatement and large-scale investment in new technology integrating production processes and abatement. The time horizon for environmental improvement, determined by a regulation authority, and certainty about what can be achieved may determine the choice between these two options.

300 Finn R. Førsund Add-on abatement requires that we make a distinction between pollutants z from the production process of (8.5), renaming them primary pollutants zP, and secondary pollutants zS actually discharged to the environment. In Førsund (2009) a production function for end-of-pipe abatement is formulated following the single-equation format of factorially determined multi-output production of having one production function for abatement of each type of residual. Here we will use a simpler formulation by focusing on a cost function in the abated amount a that is the difference between the original generation of residuals, now termed primary residuals zP, and the residuals emitted to the environment, termed secondary residuals zS. The cost function can be regarded as an index for abatement inputs and is written

c(a) = c(z P − z S ), a = z P − z S ,

c ′ > 0, c ′′ > 0 (8.9)

The abatement cost function has standard textbook properties in the abated amount.11 We will assume that the capital costs are also included in the form of yearly costs (levelized using an annuity depending on the rate of discount and the lifetime of the equipment) making capital a variable factor. The secondary residuals are assumed to be expressed in the same units as the primary residuals. The abated amounts are generally of other forms than the primary residuals. As observed in Ayres and Kneese (1969, 283) abatement does not “destroy residuals but only alter their form.” We assume that the cost of disposing of these amounts (e.g., using landfills) is included in the cost function, but that any environmental cost is zero.12 As to the materials balance principle, the abatement activity will add to the total mass of residuals if material inputs are used, but the point is that abatement means less mass of the harmful residual emitted to the environment: zS 0 z = g (x M , x S ), g x′M > 0, g x′S ≤ 0 a = zP − zS

(8.10)

P

The first term in the objective function is the area under the demand curve from 0 to y, that is, the sum of consumers’ willingness to pay for y and the revenue generated by production. The next two terms are the input and abatement costs of production. The last term is the social cost of emission of secondary pollutants. The model can straightforwardly be extended to multiple desirable and undesirable outputs, as suggested previously. When considering several units, demand functions must be adjusted according to type of demand interactions, and it must be specified whether the damage functions are unique to each unit, or the nature of interactions between damage functions must be specified if there are any (see Section 8.2). This simple model makes the fundamental trade-off between desirable output y and undesirable secondary pollutant zS tractable, allowing for optimal allocation rules both for the two types of inputs and for abatement effort. From the objective function, we see that the abatement activity does not directly influence the generation of pollutants, but indirectly by influencing the optimal solution both for primary and secondary pollution. Inserting the production functions for the good y and the pollutant zP that is the primary pollutant into the objective function, and substituting for abatement a in the abatement cost function yields

max

f ( x M , xS )

∫

u=0

p(u)du −

∑wx

j = M ,S

j

j

− c( g (x M , x S ) − z S ) − D(z S ) (8.11)

There are three endogenous variables remaining in the problem: xM, xS, and zS. The necessary first-order conditions for interior solutions are

pf x′j − w j − c ′g x′ j = 0 , j = M , S (8.12) c′ − D′ = 0

302 Finn R. Førsund There are three equations in the three endogenous variables. When these variables are determined, the optimal solutions for y, zP, and a follow directly. The two first necessary conditions tell us that for each type of factor j = M, S, the revenue of increasing factor xM marginally, and consequently increasing the desirable output, is equal to the unit cost of the factor plus the abatement cost of increasing factor xM, generating primary pollutant that is abated. For services xS negative productivities in the generation of primary pollutants imply that the last term in the first condition is positive and show the abatement cost savings of a marginal increase in a service input. The third condition tells us that at the optimal level of abatement (i.e., both primary and secondary pollutants are at their optimal levels), the marginal abatement cost should be equal to the marginal damage of the secondary pollutant. We may also have a corner solution of not using abatement at all. This will be the solution if equality between marginal abatement cost and marginal damage cannot be reached, that is, c ′(z P − z S ) > D ′(z S ) for z S ∈[0, z P ] for all zP. Abatement is too expensive. The two first conditions in (8.12) can be written

f x′j p = w j +

Marginal revenue

Unit factor cost

D ′(z S ) g x′ j ,

j = M , S (8.13)

Marginal damage cost = marginal abatement cost

At the margin, a material input generates environmental damage cost, in addition to the input price, while a service input generates a saving of damage cost by subtracting the marginal damage from the input price. The equation shows the fundamental trade-off between the value of the man-made good and the total social cost, including the environmental damage at the margin. In the case of a corner solution, the first-order condition will look like (8.13), but the marginal damage will be D ′(z P ) and both optimal desirable and undesirable outputs will be lower. To see the impact of the relative use of inputs due to different productivities in the generation of primary pollutant, the rate of substitution between a material input and a service input is

f x′M f x′S

=

w M + D ′(z S ) g x′ M wS + D ′(z S ) g x′S

(8.14)

The optimal unit price on the material input faced by the decision-maker is higher than the given market price, but the opposite is the case for the service factor because the impact g x′S on cost is negative. The optimal solution implies a relative reduced use of the material input compared with a solution without a damage function and abatement cost function. As to the substitution between material inputs, the increase in the social price follows the marginal productivity in pollutant generation, thus stimulating a shift to “greener” material inputs. It is easy to see that if a material input has a social factor price that is greater than the value of the marginal product in the desirable goods production for any value of the input, then this input should not be used. In a realistic case

Productivity Measurement and the Environment 303 of a technologically forced choice between, for example, coal and natural gas as primary energy source, the input with the lowest social factor price given the same level of marginal revenue will be chosen. The social productivity measure TFPE introduced in (8.3) for the solution to problem (8.10) and corresponding measure of productivity change are

y − D(z S ) , F (x M , x S ) + c(a) y t +1 − D ( z S , t +1 ) ∆TFPEt ,t +1 = F (x tM+1 , x St+1 ) + c(at +1 ) TFPE =

(8.15) y t − D(z S ,t ) > 1 F (x tM , x St ) + c(at )

0 λ ≥ 0 (= 0 for z S < z R )

)

(8.18)

Here λ is the shadow price on the emission constraint, that is, the gain (loss) in profit if the constraint is relaxed (tightened). If the constraint is not binding, we see from the complementarity slackness condition that the firm is not influenced by the environmental regulation and will not use any abatement. If the price of the desirable output cannot meet the unit cost including abatement cost at any level of the input j (assuming several inputs in group M), this input should not be used. Assuming that our input j is used in a positive amount and that the emission constraint is binding, the shadow price on the constraint takes the place of the marginal damage of the secondary pollutant in the optimal solution (comparing (8.12) and (8.18)). When the emission constraint is binding, the output of the desirable output in the optimal solution becomes smaller and cost becomes greater than in the private optimal solution without emission regulation because abatement has to be used and input substitution forced. The substitution effects between inputs triggered by the use of abatement resources shift the use of production inputs away from material inputs. The shift is stronger the higher the marginal productivity of the material input in question is in residuals production. The use of service inputs is increased. These substitution effects counter some of the reduction in output and profit, but these latter variables will inevitably decrease. Levels of service inputs may increase to higher levels than in the pre-regulation situation, but not enough to compensate the reduction in output and profit; if so, the firm would itself have imposed a subsidy on its use of service inputs in the pre-regulation case. The private productivity development is equal to the social productivity development in (8.15), with the important exception that the environmental damage term is absent from the numerators. Due to abatement costs in the denominator, the private productivity level goes down if output and inputs remain constant. Once regulation is introduced, the development of productivity can be either positive or negative, depending on relevant factors discussed in connection with (8.15). A new element is that

Productivity Measurement and the Environment 305 the regulated amount is subject to political changes over time. Stricter regulation is to be expected, strengthening, ceteris paribus, the case of reduced private productivity. As stated earlier, the shadow price on the emission constraint in (8.17) shows the impact on the profit of a change of the allowed emission. Marginal abatement cost has been suggested as measure of the price on pollution (Aaheim and Nyborg 1995; Pittman 1983). In the optimal solution with a binding emission constraint, we see from (8.18) that the marginal cost of abatement equals the shadow price. However, this shadow price is not in general equal to the social price on pollution. In the optimal solution of the social model (8.10), marginal damage is equal to marginal cost, so setting a price on pollutants equal to the shadow price λ in (8.18) may either undervalue or overvalue the social marginal cost, depending on the difference between the regulated amount zR and the optimal solution of zS to (8.11). It is only if the allowed amount of secondary pollutants is equal to the optimal solution of the social planner’s problem (8.11) that marginal abatement cost should equal marginal damage.

8.5. Efficiency and Environmental Regulation 8.5.1. The Porter Hypothesis On the backdrop of the huge cost that climate change may inflict on us all, and other pressing environmental problems of a more regional nature, stricter and stricter environmental regulation is introduced or is planned to be introduced in the near future. The Stern Review (Stern 2006) gives estimates of mitigating and abatement costs necessary to reach global environmental targets. A worry is that environmental regulations will set back the growth of standard goods (exclusive of environmental goods) and prevent developing countries from achieving the standard of living enjoyed in developed countries. At the micro level, a strict environmental policy imposes costs on firms and leads to lower profit and activity, according to the traditional view by most economists. This is also the case for private productivity, as pointed out in the previous section, but is not necessarily the case for productivity growth. In Porter (1991) and Porter and van der Linde (1995), a more optimistic Panglossian hypothesis, the Porter hypothesis, is put forward: strict environmental regulation may induce firms to innovate to such a degree that private profit increases, and thus strict regulation represents a win-win situation. It states that the pessimistic view stems from considering a static situation only, but that the pressure of environmental regulation induces a dynamic process of change representing retooling, process improvement, and technical change, which more than offsets the abatement costs. However, Porter and van der Linde do not present any formal mechanism supporting the cost-offset hypothesis, but refer to a few examples of successful adaptation and technical change. The Porter hypothesis

306 Finn R. Førsund and attempts to model the positive dynamics (Ambec and Barla 2002; Xepapadeas and deZeeuw 1999) and empirical studies and critique of the hypothesis (e.g., Palmer et al. 1995) are extensively reviewed in Brännlund and Lundgren (2009), Lanoie et al. (2011), and Ambec et al. (2013). The latter three references provide long lists of references to the literature on the Porter hypothesis. Notice that in the discussion of the Porter hypothesis environmental costs, the damage costs used in Sections 8.2 and 8.4, are not considered. The focus is on the possibilities of reducing emissions and the private cost of this.

8.5.2. Dynamic Effects of Environmental Regulation Porter and van der Linde (1995) suggest two different dynamic effects. First, assuming that there is inefficiency in the utilization of resources before the introduction of environmental regulation, this inefficiency is reduced or even removed after regulation has been introduced. Second, the regulation induces new technology to be developed shifting the production function outward. This is set out in Figure 8.2, which illustrates these two effects. In the space of the desirable output (q in the figure) and emissions z, the pre-regulation position of the firm is at the inefficient point C below the initial frontier production function f0(z). The efficient point A on the frontier shows the production the firm could have had corresponding to emission z0. After introducing regulation, the firm improves its efficiency and reduces the emissions down to zR and increases output from q0 to qR at point B on the initial frontier. Then there is a shift of the frontier due to innovation after introducing regulation to fR(z) where the point E is the efficient point for the level zR of the reduced emission. The firm continues to reduce emissions and increase output q, and profit Π moving toward the new frontier.14 This is a neat possible illustration of the story told in Porter and van den Linde (1995) of the increased efficiency effect and the shift in technology effect, but it does not explain the mechanisms behind the moves. Production, q E

q1 Π1 qR

A

q = fR(Z) q = f0(Z)

D

B

ΠR q0

C

Π0

zR

z0

emissions, z

Figure 8.2. The Porter hypothesis. Source: Brännlund and Lundgren (2009, Panel (b), 83).

Productivity Measurement and the Environment 307 Establishing credible dynamic mechanisms has proven to be rather difficult. Regarding inefficiency, this difficulty is underlined by the fact that there are hardly any dynamic mechanisms in the huge body of specialized literature on inefficiency explaining the reduction of inefficiency. Estimating efficiency scores based on parametric and nonparametric production functions, and indices for productivity change of the Malmquist type using technical efficiency scores, may be said to be no more than more or less advanced descriptions of implications of observed data on outputs and inputs and do not represent explanations of inefficiency. There is a general methodological problem here, namely the difficulty in explaining inefficiency based on rational decisions. One type of explanation mentioned in the early efficiency literature (see, e.g., Førsund 2010 for a review) is that the inputs may be heterogeneous (Farrell 1957 mentions capital), implying that the estimated inefficiencies are biased. Stigler (1976) goes so far, in reviewing the Leibenstein (1966) concept of X-efficiency, to state that there is no inefficiency in competitive industries, just inhomogeneous inputs, including not only capital and labor, but also management (Charnes et al. 1978 point to the latter factor as a cause of inefficiency). Inefficiency can be rationalized if the technology in production capital and equipment is embodied (Førsund and Hjalmarsson 1974). Then, if a firm having old technology is compared with one having a new technology, it may appear as inefficient. However, with long-lived equipment, this type of inefficiency may well serve an objective of maximizing the present value of profit (Førsund 2010). The standard critique of the Porter hypothesis builds on the following questions: Why haven’t good ideas already been utilized? And why should the firms wait until environmental regulation comes along? If profitable opportunities to reduce emissions of pollutants exist, profit-maximizing firms would already be taking advantage of them (Ambec et al. 2013). Apart from pointing to limited information about possibilities and limited resources, support for the dynamics of the Porter hypothesis can be found in the theory of induced innovations of Hicks (1932). Porter and van der Linde (1995) talk about the pressure put on firms to comply with environmental regulations. The pressure in the theory of Hicks is represented by costs of inputs (i.e., resources are devoted to R&D in order to reduce costs). However, the development of technology may have deeper roots than mere cost savings. There is a path from basic research, without any cost considerations firms may have about usefulness, to applied research and to innovations within firms themselves or within equipment-producing industries. However, it is not easy to establish a dynamic mechanism for events that are stochastic by nature.15 We can distinguish between three types of reactions to new environmental regulation by a firm: (1) short- run measures like better process control, small- scale re- engineering, introducing more internal recycling of waste, leading to improved efficiency of utilizing material inputs and thereby reducing pollutants, (2) medium-term measures like investing in end-of-pipe technologies,

308 Finn R. Førsund (3) long- term measures involving developing new technologies for the main processes saving material inputs and reducing waste products, and even developing waste products into salable products. Technologies used in material- processing industries often have embodied specific technologies so technical change can only come by investing in new technologies and scrapping old machinery and equipment. It may be the case that a choice has to be made between investing in short-term measures or long-term measures. However, going for short-term measures may lead to less emphasis on developing new process technologies and thus may not be a socially optimal decision. Uncertainty about future regulation may imply favoring short-term measures. The requirement for technical change to lead to increased private profit can be expressed using the relations (8.8). In addition, there may be technical change in the abatement function that we can express using the cost function (8.9): c t2 (z P − z S ) < c t1 (z P − z S ),

t 2 > t1 (8.19)

The periods t1 and t2 may not necessarily be consecutive years, but may span a period more reasonable for technical change to take place. We will therefore consider capital as one of the factors, but without formulating any investment problem that should have to be done in a realistic setting. Let us assume that the environmental regulator set an upper limit on allowable emission, as in model (8.16), but let the firm itself determine how to obtain this level. We assume that the regulation starts in period t1 and that the regulated amount is the same and binding in both periods. It is reasonable that it takes some time before the regulatory pressure leads to new technology being available, so the difference between t1 and t2 may be several years (Lanoie et al. 2008). The product price and the factor prices, respectively, are the same in both periods by assumption. The condition for the pressure to lead to increased profit is then py t2 −

∑wx

j = M ,S

j

t2 j

− c t2 (ztP2 − z R ) > py t1 −

∑wx

j = M ,S

j

t1 j

− c t1 (ztP1 − z R ) ⇒

p( f t2 (x tM2 , x St2 ) − f t1 (x tM1 , x St1 )) + ∑ j = M ,S w j (x tj1 − x tj2 ) Output effect

(8.20)

Input effect

+ (c t1 ( g t1 (x tM1 , x St1 ) − z R ) − c t2 ( g t2 (x tM2 , x St2 ) − z R )) > 0 Abatement cost effect

In period t1 the initial output effect is negative, but input costs probably go down, but then abatement costs contribute to reducing the profit. There are three effects of technical change in period t2: the change of the amount of output, the change in the amounts of inputs, and the change in the abatement costs. As to the latter effect, we also have a negative cost consequence due to the reduction in primary pollutant following technical change. Technical change satisfying the condition (8.20) will have a positive output

Productivity Measurement and the Environment 309 effect, an uncertain input effect, and a reduction in the abatement costs because of the two technology shifts in the second period. If the effect on profit is positive, then private productivity will also increase over time. To keep up a pressure on firms that have managed to comply with the regulation, it may be necessary to have more stringent regulation over time to keep the dynamic process going.

8.5.3. Regulation Using Economic Incentives The type of regulation introduced can play a decisive role for the reactions of firms. Porter and van der Linde (1995) base their dynamic innovation process on well-designed environmental regulations based on economic instruments. They go strongly against using a command-and-control approach that demands that a specific technology be used. This will stifle any innovation activity of the firms exploring other solutions and may not lead to any further development of technology if the command of a specific technology to be used satisfies the required reduction in emissions. The technology imposed by the regulator is usually of the type “best available technology” (BAT) or “best available technology not entailing excessive cost” (BATNEEC). Installing end-of- pipe abatement may be cheaper and less complicated than demanding new basic process technologies, so the regulator may be biased toward imposing end-of-pipe solutions without considering the socially optimal solution in a longer time perspective. However, regulators may also give firms a certain number of years before complying if changing the process technology is the best solution. Thus, existing firms may enjoy a positive quasi-rent from “dirty” processes, biasing against new firms entering the market and having to invest in new, less polluting technology. Jaffe and Stavins (1995) study empirically the importance of dynamic incentives of regulation for technology diffusion and find economic instruments more effective than direct regulations. The seminal economic instrument is to introduce a Pigou tax on the emitted pollutant (Pigou 1920). Using our model (8.5) extended with (8.9), the environmental regulator uses a tax on secondary pollution as the instrument giving incentive to react in both the short and the long run. Regarding the unit as a firm that maximizes profit, facing competitive markets for both output and inputs, introducing a Pigouvian tax t on secondary pollutants yields the following optimization problem:

max pf (x M , x S ) −

∑wx

j = M ,S

j

j

− c( g (x M , x S ) − z S ) − tz S (8.21)

The production functions for output and pollutant from (8.5) are inserted for output and primary pollutant, respectively, and a flat unit tax t is imposed on the secondary pollutant. We assume that the firm has access to an end-of-pipe abatement technology, and we regard capital as a variable factor in the cost function as explained in connection with (8.9).

310 Finn R. Førsund The necessary first-order conditions, assuming that abatement will be used, are

pf x′j − w j − c ′g x′ j = 0 , j = M , S (8.22) c′ − t = 0

The optimal firm solution will be a function of the exogenous tax rate t. Going back to (8.12), we see that the firm solution will conform with the social solution if the tax rate is set equal to the marginal environmental damage in the latter optimal solution. The choice between using and not using abatement is decided by the last condition in (8.22): if the marginal cost is higher than the unit tax for all levels of abatement, then it is optimal to set abatement equal to zero and emitting the primary pollutant with the tax levied on it, yielding the condition pf x′j − w j − tg x′ j = 0. Because the tax t is exogenous, we can solve these two equations for the inputs xM and xS. Then the solutions for the outputs y and zP follow directly. With positive abatement, the optimal solution to problem (8.21) implies that the marginal abatement cost is set equal to the tax. We then see that we get the same solution for the inputs and the desirable and undesirable outputs in the two cases, but without the abatement being optimal to use, the emission to the environment zP is higher than the secondary pollutant zS in the case of the abatement facility being profitable to use. Going back to the discussion in connection with Equation (8.12), without abatement being socially optimal to use, the tax rate must be set at a higher level to realize a certain goal for environmental pollution. The relative use of factors will be changed in the same way as illustrated by (8.14), and output will be scaled down, compared with the profit-maximizing solution, without a charge levied on the pollutant and no use of abatement. The results are based on static technologies; substitution takes place within the pre- regulation technology. According to the Porter hypothesis, the tax introduces a pressure on the firms to try to reduce the tax “punishment” in the short run by diminishing any existing polluting waste, and in a longer run adopt end-of-pipe solutions that reduce the tax burden, and develop new technologies using less raw materials that generate pollutants per unit of output. The impact on productivity and productivity change is similar to what is described in the previous subsection. However, a new element has entered the numerator: the tax paid on the secondary pollution is deducted from the private profit. Again, the static productivity is reduced, but the productivity change may be both negative and positive according to the strength of the different factors, as discussed earlier. Now the tax amount is a new influence on the productivity change if it increases over time. Direct regulation in the form of an upper limit on pollutant emitted to the environment should also work as a pressure to adapt and adjust the technology and to innovate basic process technologies. We see from the necessary first-order conditions (8.12) for optimal solution of model (8.11) that these conditions are of the same form as in (8.22). However, there is a difference as to the strength of the pressure because the regulated amount zR is not necessarily the optimal amount zS*, and in the case of using a tax this has to be paid in addition to the abatement costs, thus reducing the private profit

Productivity Measurement and the Environment 311 below the socially optimal one.16 With the direct regulation of this type, the firm has the same freedom as under a tax regime to develop ways of reducing the use of pollution- generating raw materials and to develop technologies either of the end-of-pipe type or of process type doing the same. For the regulator, the difference is that a known maximum amount zR is emitted to the environment imposing such a limit, while with the tax solution the regulator may be uncertain about what will actually be realized as the emission (assuming that the regulator does not have perfect information about the firm’s technology).

8.5.4. Cap and Trade It is well known from the environmental economics literature that “cap and trade” can combine both types of instruments for a situation in which a number of sources emit the same type of pollutant or pollutants that have the same effect on the environment independent of the location of the source (e.g., greenhouse gases and their effects on global temperature). By imposing an upper limit on the total amount of emission from all firms and then introducing quotas for each firm made tradeable, the equilibrium price of quotas will act as if a tax on pollutants is introduced, as in problem (8.21). The optimization problem for firm i (i =1, . . . , N) is now

max pfi (xiM , xiS ) −

∑wx

j = M ,S

j

ij

− ci ( g i (xiM , xiS ) − ziS ) + pQ (ziQ − ziS ) (8.23)

We have three types of endogenous variables: xM, xS, and zS. The quota for firm i is ziQ and we assume that ziQ < ziS0 for all firms, where ziS0 is the optimal level of secondary residuals for unit i before the cap and trade regime is introduced. The total emission then goes down due to the caps. We assume that the quota price pQ is clearing the quota market. The emission ziS is the optimal emission under the cap and trade regulation. The necessary first-order conditions are

pfix′j − w j − ci′g ix′ j = 0 , j = M , S (8.24) ci′ − pQ = 0, i = 1,..., N

Having obtained the solution for the endogenous variables xM, xS, and zS, the solutions for y and zP can be obtained. The first condition shows that use of material inputs will be reduced; the unit cost will go up, but the use of service inputs will go up; their unit cost is reduced. It is reasonable to assume that the firm will react with a cutback in production, thus reducing revenue and profit. As to buying or selling quotas, the crucial condition is the second one saying that marginal abatement cost should be equal to the quota price. Since the quota price is the same for all firms, the marginal abatement costs will also be the same across firms. The firms that abate, but have optimal emissions that exceed the quota, ziS > ziQ, have to buy additional quotas. It is possible that marginal abatement cost is higher than the quota price for zero abatement, then ziP – ziQ has to be bought on

312 Finn R. Førsund the quota market. The sellers will be firms for which it is profitable to abate so much that ziS < ziQ. It is formally possible that marginal abatement cost is lower than the quota price for maximal abatement, that is, a = ziP and ziS is zero, although observations typically show that this is not the case. We see that all firms that abate will face the same marginal abatement cost equal to the clearing price pQ in the market. Comparing this solution to the solution (8.24) to problem (8.23) with a tax on the emission we have that, provided that the total restriction on emissions is the same as the sum of optimal solutions for emissions when using a tax, the market clearing price is equal to the tax. The advantage with cap and trade is that it combines three features; the certainty of regulated emissions, provides perfect compliance, and gives economic incentives or pressures to improve efficiency and develop new technologies that reduce emissions. An additional advantage is that the regulator can tighten the total amount of quotas over time as a single decision. However, a tax on emissions and cap-and-trade may have different incentives in a dynamic setting. A tax represents a predictable form of regulation for a firm, assuming that it is not changed so frequently. But the price on emission permits may fluctuate quite a lot and represents an unpredictable factor for investment decisions related to environmental regulations. Behind fluctuations are the business cycle, technology change, competition from firms in countries not regulating carbon emissions, and the thermal electricity generators’ competition from renewable energy with its fluctuations of production. The incentives of the two different instruments may give incentives to investment in different technologies, presumably more short-term considerations in the form of end- of-pipe technologies than long-term more costly and fundamental process innovations due to uncertainties about the price. The fluctuations are illustrated in Figure 8.3, showing the price development for carbon certificates17 within the European Union

35 30

EUR

25 20 15 10 5

14

13

20

12

20

11

20

10

20

09

20

08

20

07

20

06

20

20

20

05

0

Figure 8.3. Volatility of the EU Emission Trading System (ETS) carbon certificate market (2005–2014). Source: Alexandru P. Luta. The current state of the EU ETS. The Sanbag Climate Campaign, July 22, 2014. https://www.google.no/webhp?gws_rd=ssl#q=The+current+state+of+the+EU+ETS.

Productivity Measurement and the Environment 313 Emission Trading System (ETS) for 10 years. The preceding theoretical model of the quota market links marginal abatement costs to the quota price. However, estimates of abatement costs of participating firms fluctuate much less. The EU ETS has had three phases, the first phase covering the period 2005–2007, the second phase 2008–2012, and the third phase 2013–2020. The scheme started out with free quotas for about the total emissions. A positive price in spite of the total free quota is due to mismatch between firm-specific quotas and actual emissions. The fall in price during 2007 is due to earlier trades leading to a closer match between quotas and emissions, and the fact that quotas bought earlier but saved to be used later could not be transferred to Phase II. In Phase II the total quota was reduced somewhat below the emissions, and more countries joined the system. The price started at levels similar to the start of Phase I, but then the economic recession caused by the financial crisis started to bite and reduced the output of firms within the system and the quota price plummeted. The continuing slide and low level, which has continued to date, is due to structural adjustments by firms in the participating sector, such as fuel substitution of thermal electricity generators and the increased share of renewables in electricity production.

8.5.4. Empirical Studies of the Porter Hypothesis In the substantial literature on the Porter hypothesis, it is difficult to find general theoretical arguments on which to build mechanisms that give the results of the hypothesis. Therefore, the validity of the hypothesis may be regarded as an empirical question. There is also a rich literature trying to do this, but there is a high ratio of speculation and anecdotes to real evidence (Jaffe et al. 1995; Lanoie et al. 2008), so it is difficult to generalize the obtained results. We will not give a detailed account of empirical studies here, but will focus on the main findings.18 A common approach is to study separately three aspects of the Porter hypothesis; the weak version, the narrow version, and the strong version (Lanoie et al. 2011). The weak version posits that environmental regulation will stimulate innovations and technology choice to make production “greener,” as measured by R&D, investment in capital, and patents. The narrow version states that flexible environmental regulation gives firms stronger incentive to innovate than command-and-control regulation prescribing specific technologies. The strong hypothesis states that properly designed regulation may induce innovations that more than compensates for the cost of compliance measured by firms’ performance such as productivity and costs. The main findings are that the weak hypothesis has some empirical support, as is the case for the narrow hypothesis, but that there is not much evidence for the strong hypothesis of increased profit due to regulatory-induced innovations. In the studies of innovations, attention is paid to time lags between the introduction of environmental regulation and innovations. There is a transition period in most cases. Contemporaneous impact on economic performance may be negative, but can turn out positive when using lags between regulatory variables and effects (Lanoie et al. 2008).

314 Finn R. Førsund There is a data problem in identifying the environmental cost element of process- changing investments that may have been done without regulation. The lack of accurate accounting of regulation cost that allocates the parts attributable to environmental regulation and to process changes undertaken without regulation is becoming a problem because there is an increase in process changes and product reformulations rather than installation of end-of-pipe equipment.

8.6. Productivity Measures Based on Inefficiency An interpretation of the Porter hypothesis is that existing inefficiency (before the introduction of environmental regulation) is reduced or even eliminated as low-hanging fruits in the short run, and in the long run inefficiency seems to disappear completely, at least it is not mentioned in connection with long run. However, there is a body of literature on productivity that assumes that inefficiency may be persistent over time, the main reason being that innovation and new technology shift the frontier production function upward over time, and consequently make it harder for firms to become as efficient as current best practice. This strand of research is commonly based on the assumption that prices are not available, and the estimation of the efficiency scores is mainly based on a nonparametric frontier (for reviews of this literature, see Arabi et al. 2015; Arabi et al. 2016; Dakpo et al. 2016).

8.6.1. Single-Equation Models Pittman (1983) extended the approach of Caves et al. (1982a) of measuring productivity, introducing translog multilateral productivity indices, to include undesirable outputs or pollutants. Shadow prices, mainly based on abatement costs, on undesirable outputs were used in calculating the output index based on revenue shares and an input index based on cost shares. Färe et al. (1989) were not satisfied with the approach of Pittman (1983), pointing to difficulties in determining the prices of undesirable outputs. They introduced a hyperbolic efficiency measure based on estimating Farrell (1957) efficiency measures, using only quantitative information about inputs and outputs, specifying a nonparametric frontier function. This was done in order to credit producers for “their provision of desirable outputs and penalize them for their provision of undesirable outputs” (90). Undesirable outputs z were introduced in addition to the standard desirable outputs y. The starting point is to introduce a technology using the output possibility set

P (x ) = {( y , z ) : x ≥ 0 can produce y ≥ 0, z ≥ 0} (8.25)

Productivity Measurement and the Environment 315 where y, z, and x are vectors. In order to give the technology some structure, it is common to start with properties that the set should fulfill. In order to be as general as possible, the set P(x) is assumed to be convex, closed, and bounded, and is monotonic in y, z, and x; both y and x are strongly disposable, but there is a question about the disposability prop erty of the undesirable output that will be addressed in the following. If we assume that also z is strongly disposable, an output-oriented distance function (Shephard 1953) can be used to characterize the technology:

DO ( y , z , x ) = min{θ : ( y / θ, z / θ) ∈P (x )} ∈(0, 1]. (8.26)

The point ( y / θ, z / θ) on the border of the set P(x) is the largest feasible radial extension of the observed output vectors y and z for a given x vector. The distance function gives the Farrell output-oriented efficiency score for the observation (y, z) for a given x. The relation DO ( y , z , x ) = 1 gives the implicit relation between outputs for a given vector x along the efficient border of the set. This means that we have a trade-off between outputs, that is, inputs can be reallocated to any desired mix of outputs (the assortment is maximal, in the language of Frisch 1965). If we have both desirable and undesirable outputs, this implies that resources can be reallocated away from all undesirable outputs and thus we do not have any problem with undesirable outputs. The trade-off between desirable and undesirable output y and z, respectively, must therefore be restricted in some way for the technology to lead to sensible results (Färe et al. 2013, 110; Førsund 2009). This is accomplished in Färe et al. (1989) and the literature that followed by imposing weak disposability (introduced in Shephard 1970) for y and z together:

( y , z ) ∈P (x ) and 0 ≤ θ ≤ 1 imply (θ y , θz ) ∈P (x ). (8.27)

Furthermore, it is assumed that y cannot be produced without z also being produced; this null jointness is defined as: if ( y , z ) ∈P (x ), and z = 0, then y = 0. However, although the intention of measuring efficiency taking into account generation of pollutants is a good one, the solution using the assumption of weak disposability is not (see critique in Førsund 2009, 2018; Murty et al. 2012; Murty and Russell 2018). As pointed out in Subsection 8.3.2, the generation of pollutants stems from the use of material inputs. Then the materials balance accounts for where the mass content of the raw materials ends up; incorporated in the desirable output or as pollutants. There is physically no possibility of a technical trade-off between a desirable output and a pollutant for given amounts of inputs.19 But such a possibility is not only allowed by the formulation (8.27), but is crucial for the modeling, using a radial or directional distance function as the single-equation model (see the many illustrations in the weak-disposability literature of such a trade-off, starting with Shephard 1970, 188). One should be aware of the fact that a production possibility set is a very general statement to make about production possibilities until more structure is introduced. Moreover, there is nothing in a general specification saying that such a set can only be described by a single equation such as a distance function. The modelling of joint production may involve a number of equations in order to capture some essential feature of

316 Finn R. Førsund the technology, as demonstrated in Frisch (1965). The technology set may be the intersection of many subsets. This seems to be overlooked in the single-equation literature. The point is how to model the essential connection between desirable and undesirable outputs on one hand, and the inputs on the other. The simultaneous generation of both desirable and undesirable outputs transforming inputs (including at least one material one) is the essential feature. This implies that to look at a relation between the desirable and undesirable outputs for given inputs is futile; there is no transformation possibility here due to the materials balance principle and the assumption of an efficient utilization of all inputs in obtaining the desirable output (Førsund 2018). The problem dealing with both desirable and undesirable goods when prices on the latter do not exist was addressed in Färe et al. (1989), calculating a hyperbolic index of the type

H ( y , z , x ) = max{λ : (λ y , λ −1z ) ∈P (x )} (8.28)

This idea of asymmetric treatment has in the last decade been developed by estimating efficiency scores using directional distance function (Arabi et al. 2015; Arabi et al. 2016; Chung et al. 1997). The directional distance function is based on choosing a direction from an inefficient observation (adding or subtracting from the observed values) to the frontier that is not necessarily radial. For an output-oriented efficiency score, the value of the output point on the frontier is measured adding to the observed output the product of a common expansion factor and the desirable outputs, and in the case of an undesirable product to subtract the product of the expansion factor and the value of the observation of an undesirable output. The directional output distance function is defined by

Do (x , y , z ; g y , − g z ) = sup[β : y + β g y , z − β g z ] ∈P (x ). (8.29)

The directions gy and gz are usually chosen as (y, z). The minimal value of the directional distance function is zero when the unit in question is on the frontier, and it is restricted to be greater than or equal to 0. The expansion factors for desirable outputs with directions (y, z) are (1+β) and the contraction factor for undesirable outputs is (1 – β). However, apart from the problem that the directional distance function goes against the material balance and efficiency of production relations, as pointed out earlier, there is the special feature of imposing the same expansion of goods as the contraction of bads. This seems to be rather arbitrary. “Punishing” the generation of pollutants with the same factor as “rewarding” desirable goods is just one of many ways of implicitly weighting the two types of outputs. The fundamental definition of environmental damage we are seeking may be far from the interpretation of efficiency measures based on directional distance functions. It may be useful for the reader to see how a Malmquist type of productivity-change index, termed Malmquist-Luenberger index (ML) in Chung et al. (1997), is set up using directional output distance functions for period t and t +1:20

1 + Dot (x t , y t , z t ; y t , − z t ) ML = . (8.30) 1 + Dot (x t +1 , y t +1 , z t +1 ; y t +1 , − z t +1 ) t +1 t

Productivity Measurement and the Environment 317 MLtt+1 > 1 implies Dot > Dot+1, that is, that period t efficiency for the unit in question is less than the efficiency in period t +1, resulting in an increase in productivity. The crucial point for environmental policy is a trade-off between values of goods and pollutants. However, this must be based on a trade-off between man-made desirable goods and jointly generated pollutants measured in the same unit. It is difficult to see that this can be captured by defining productivity as changes in directional distance functions with a fixed goods expansion factor and the same factor as a contraction factor of pollutants. Productivity measures consistent with real trade-offs are inevitably required by regulatory activities, carefully balancing benefits and costs (Jaffe et al. 1995). Another weakness in the way that single-equation models based on distance function have been used is the lack of explicit modelling of abatement activities, as is done in the model in Subsection 8.3.4. Measuring abatement in the form of only lost output is taking into account only one of the ways to reduce the generation of pollutants mentioned in Subsection 8.3.3. It is standard in the environmental economics literature to model abatement as end-of-pipe, as used in Subsection 8.3.4 (see Førsund 2009, 2018). Abatement is introduced in Färe et al. (2013) as a separate activity, using the primary pollutant, shares of resources and a share of the output from the primary process as inputs producing the secondary pollutant. However, generating the two types of outputs, weak disposability is still maintained as an assumption, so this approach to estimate efficiency including abatement is problematic to accept.

8.6.2. The Multi-Equation Model and Inefficiency The model of factorially determined multi-output production used in Section 8.3 highlights insights from environmental economics. This model can easily be generalized to include inefficient operations. The simplest way is to use the three relations (8.5) and (8.9) depicting frontier functions, obeying the materials balance (8.7), by expanding the functions involved into sets:

y ≤ f (x M , x S ) ⇒ T1 = {( y , z , x ) : y ≤ f (x M , x S )} z P ≥ g (x M , x S ) ⇒ T2 = {( y , z , x ) : z ≥ g (x M , x S )} (8.31) c o ≥ c(a) = c(z P − z S )

Following Murty et al. (2012), the production possibility set T is the intersection of the two first sets; T = T1 ∩ T2 . The observed cost for abatement a is co. The production and cost frontiers are on the right-hand side of the inequalities. Strict inequalities open for inefficient operations, while equations holding with equalities mean efficient operations. (Recall from Subsection 8.3.5 that introducing abatement activity makes it necessary to operate with two types of emission; primary pollutants and secondary pollutants.) We can distinguish between three types of efficiency measures: (1) Efficiency in producing desirable outputs: E y = y obs / f (x M , x S ) ∈(0,1] (2) Efficiency in generating pollutants: Ez = g (x M , x S ) / z P obs ∈(0,1] (3) Efficiency in abatement cost: Ec = c(a) / c o ∈(0,1] .

318 Finn R. Førsund The index obs indicates the observations of the three variables, and in addition the inputs xM, xS, and abatement a are also the observed ones. Productivity development can be calculated using, for example, a Malmquist productivity change index21 based on efficiency calculations for each of the three frontiers of the equations (8.31). These equations are independent; y is not depending on z and vice versa, and abatement is an independent stage. From a policy point of view, this is the preferred information. It does not seem to make sense from a policy point of view to try to estimate a total or aggregated efficiency and productivity. Keeping the developments separate also underlines the point that to see the trade-off between desirable and undesirable goods is only something that can be achieved if the effects of pollution in the environment are made comparable with man-made goods by measuring each type of output in the same unit. It does not make much sense to try to capture a trade-off by forcing a numerical trade-off not based on real evaluation of values.

8.7. Conclusions The main question has been how to include considerations of the natural environment when measuring productivity of production and its change over time. The natural environment provides us with services of various kinds, influencing our health and well- being in a profound way. However, many of these services are not traded in markets and also have a public good character. It may therefore be futile to search for prices of these services. What can be done for monetized evaluation is to estimate the damages inflicted on the environment as evaluated by consumers of environmental services. Concentrating on how to deal with pollution from production using material resources, estimates of damages constituting the damage function are based on the value of man- made desirable goods that we are willing to forsake in order to reduce the pollution from residuals that are generated jointly with the desirable outputs. The guiding principle for formulating a model describing joint generation of intended and unintended outputs has been the materials balance principle. Matter cannot be created, nor can it disappear, so the matter of raw materials must either be part of the desirable outputs or end up as residuals discharged to the environment. A multi- equation model is then most suitable to capture the simultaneous generation of desirable and undesirable outputs. The choice was the factorially determined multi-output model having separate production functions for desirable outputs and residuals. This model does not allow any functional relationship between the desirable output and the undesirable one for given amounts of inputs; just a single point in output space is determined, thus obeying the materials balance. Productivity is measured as the ratio between aggregated outputs and inputs. Outputs are either value added or gross output, and volume indices have to be constructed in the multi-output case. Inputs can be just labor, or may comprise all types in the case of multifactor productivity pursued here. Environmental concerns are then entered

Productivity Measurement and the Environment 319 by forming the social gross output, subtracting damage costs from gross “economic” outputs. For constant outputs and inputs, static productivity then has to go down. However, once the adjustment is done, productivity change may be both positive or negative, depending on the time path of the different components. Within a general equilibrium framework, resources are channeled to pollution-reducing activities, thus leading to a general reduction of man-made goods, but it is still possible for productivity change incorporating damage cost to show an increase. However, capital accumulation may go down in general and may be a brake on productivity growth if not compensated sufficiently by increase in environmental qualities (i.e., a reduction in damage cost). Generation of pollutants can be reduced in this model by substitution between inputs (e.g., fuel substitution, and using non-material inputs instead of material ones), keeping desirable output constant, and reducing the amount of desirable outputs. Another way of abatement is changing technology, which usually is possible in the long run only, but there is also an option of installing end-of-pipe facilities that can be done within a shorter time horizon. Such a facility is added to the factorially determined multi-output model (8.5). As for productivity, abatement costs represent additional inputs reducing static productivity, but again productivity change may go either way; this is an empirical question also influenced by technical change in the long run. The Porter hypothesis postulates environmental regulation based on the “right” incentives to be a win-win situation; regulation puts pressure on firms to reduce current inefficiency and to innovate to such an extent that profit and productivity increase. However, there is not much systematic evidence for such a Panglossian view. No proper formulations of formal dynamic mechanisms have been put forward. In the literature on inefficiency, environmental considerations have also been developed. The bulk of inefficiency modelling has used a single-equation model based on a special assumption called weak disposability, linking desirable and undesirable outputs to move proportionally at the efficient border of the production possibility set. Weak disposability (together with null jointness) blocks the reduction of pollution to zero, keeping positive amounts of desirable outputs, but no economic or engineering explanations have been put forward backing up these assumptions. The physical background for the generation of residuals when using raw materials is not commented upon at all, and the materials balance is neglected. Indeed, the weak disposability goes counter to the materials balance principle because it is based on a transformation possibility (assorted production) between the desirable and undesirable outputs for given amount of resources. However, it is impossible to decrease both desirable and undesirable outputs simultaneously for given resources remaining on the efficient part of the frontier function (or the efficient border of the production possibility set). Saying that resources are just made idle goes against the definition of an efficient border of the set, and telling a story that resources in production are reduced because they are reallocated to abatement without modelling such activity is not acceptable. Calculations of productivity and its change have been done applying directional distance function within the weak disposability framework and using a Malmquist type of productivity measure. To make productivity of desirable goods comparable with

320 Finn R. Førsund pollutants, the expansion factor for undesirable outputs is restricted to be the same as for desirable outputs, but with a negative sign. But this is just an arbitrary way of comparing productivity growth, having nothing to do with a proper evaluation of pollutants and desirable goods; the task is to make desirable man-made goods comparable with environmental services in value terms. The strong conclusion of this chapter is that the use of single-equation models based on the assumption of weak disposability should be at the end of its road. However, it should be stressed that it is not the assumption of weak disposability that is the main problem: the point is that a single-equation model assuming a trade-off between desirable and undesirable outputs is not compatible with the materials balance and the efficiency requirement of the (implicit) production relations at the efficient border of the production possibility set. Multi-equation models capturing essential mechanisms behind simultaneous generation of pollutants and desirable goods should be developed. The solution to the problem of incorporating pollution into productivity measures has been based on introducing an environmental damage function. However, although a neat theoretical solution, it may not be so helpful for practical calculation. Further building up of satellite accounts is needed, tracking physical changes in various types of natural environments and stocks of resources, both biological renewable ones and nonrenewable resources. An alternative to estimating damage costs is to impose physical limits on the discharge of pollutants if the physical damages are known well enough. As demonstrated with cap-and-trade schemes, setting physical limits implies that mitigation costs are revealed, thus making a trade-off between man-made goods and pollutants possible.

Notes 1. See Mishan (1971) for a review of the earlier externalities literature, and Fisher and Peterson (1976) and Cropper and Oates (1992) for reviews of the literature covering the 1970s and 1980s. 2. See Chapters 2 (Balk), 3 (Eldridge, Sparks, and Stewart), and 4 (Russell) in this volume for more detailed accounts. 3. See Chapter 20 (Jorgenson) in this volume on the international KLEMS project. 4. The Maximum principle was published in 1956 in Russian and an English version in 1962, and the mathematical technique was applied to economic problems in the 1960s—a fast and widespread diffusion of an intellectual invention. 5. The adaptation of the Frisch multi-equation model to environmental pollution is based on the models presented in Førsund (2009, 2018). 6. In the case of inefficiency that will be introduced in Section 8.6, the functions f(.) and g(.) represent the efficient border of the production possibility sets and are called frontier functions in our setting. 7. Cf. the famous chocolate production example in Frisch (1935) discussed in Førsund (1999), of short-run substitution between labor and cocoa fat due to more intensive recycling of chocolate with moulding defects the more labor that is employed. Moulding defects decrease with a higher proportion of cocoa fat. Stricter quality control using more labor reduces the share of defective products, utilizing the raw material more efficiently.

Productivity Measurement and the Environment 321 8. One or more service inputs may also be essential, however, the point is that residuals are in general an unavoidable feature using material inputs in production. Although y = f (x M , 0) = 0 we may have z = g (x M , 0) > 0 ; for example, a fully automated thermal electricity-generating plant running in a spinning mode. 9. This theme is developed in the ecological economics literature; see, e.g., Pethig (2006). 10. Retooling or re-engineering processes, and recycling internally more waste materials, may be done within a reasonably short period, and small-scale investments like heat exchangers to recapture waste heat can reduce the amount of residuals for constant primary energy, and thus increase production (Martin 1986). 11. Factor prices of abatement inputs are not introduced because the employment of individual abatement inputs will not be studied (see Førsund 2009, 2018 for such studies). 12. This may not always be the case; there may be harmful run-offs from landfills, and the incineration of waste may create harmful gases, but these factors can in principle be entered into the analysis at some cost of extending the equation system (Førsund 2009). 13. At an aggregated or macro level, the resources used on abatement mean fewer resources to other activities, but at the micro level of a single firm operating in competitive markets, this is not a concern in a partial equilibrium analysis. 14. As opposed to the model (8.5), the pollutant is functioning as an input. This can be the case if there is a fixed relationship between the real input and the pollutant in a single desirable output and a single input and pollutant world. The pollutant is then a shadow factor to the material input, using the terminology of Frisch (1965, 22). 15. Innovations may also be directed to develop new products. As mentioned before, this is not addressed in this chapter. 16. This is a well-known result in textbooks on environmental economics (Førsund and Strøm 1988). 17. Note that it is carbon that is the physical basis in the market, not CO2. This means that γz in (8.7) is the basis. 18. For detailed accounts, see Brännlund and Lundgren (2009), who give extensive tables of references with information on purpose and method, data, and results; Lanoie et al. (2011), who used a data set for approximately 4,200 facilities collected from seven OECD countries by postal survey in 2003; and Ambec et al. (2013), who also give an account of many empirical studies. 19. This is extensively discussed in Førsund (2009, 2018). 20. In Chung et al. (1997), the ML index is computed as the geometric mean using the technology from period t and period t + 1, respectively, as the base technology, following the practice in the literature set by Färe et al. (1992, 1994). However, we follow here the original Caves et al. (1982b) using the technology from period t only as the base technology. Notice that there is a problem of comparability over time if the directions for each period are taken as the observations. 21. The original popular Caves et al. (1982b) Malmquist productivity change index is not a true total factor productivity index. However, the total factor productivity index developed in Bjurek (1996) is a productivity change index, defined as a Malmquist volume index for outputs in the numerator and a Malmquist input index in the denominator and is therefore a true total factor productivity index. This index is also called the Hicks-Moorsteen index, following a suggestion in Diewert (1992).

322 Finn R. Førsund

References Aaheim, A., and K. Nyborg. 1995. “On the Interpretation and Applicability of a “Green National Product.” Review of Income and Wealth 41(1): 57–7 1. Ambec, S., and P. Barla. 2002. “A Theoretical Foundation of the Porter Hypothesis.” Economics Letters 75(3): 355–360. Ambec, S., M. A. Coheny, S. Elgiez, and P. Lanoie. 2013. “The Porter Hypothesis at 20: Can Environmental Regulation Enhance Innovation and Competitiveness?” Review of Environmental Economics and Policy 7(1): 2–22. Arabi, B., S. Munisamy, and A. Emrouznejad. 2015. “A New Slacks-Based Measure of Malmquist– Luenberger Index in the Presence of Undesirable Outputs.” Omega 51(March): 29–37. Arabi, B., M. S. Doraisamy, A. Emrouznejad, and A. Khoshroo. 2017. “Eco- Efficiency Measurement and Material Balance Principle: An Application in Power Plants Malmquist Luenberger Index.” Annals of Operations Research 255(1–2): 221–239. doi: 10.1007/ s10479-015-1970-x. Asheim, G. B. 2000. “Green Accounting: Why and How?” Environment and Development Economics 5(1): 25–48. Ayres, R. U., and A. V. Kneese. 1969. “Production, Consumption and Externalities.” American Economic Review 59(7): 282–297. Bjurek, H. 1996. “The Malmquist Total Factor Productivity Index.” Scandinavian Journal of Economics 98(2): 303–313. Brännlund, R., and T. Lundgren. 2009. “Environmental Policy Without Cost? A Review of the Porter Hypothesis.” International Review of Environmental and Resource Economics 3(1): 75–117. Caves, D. W., L. R. Christensen, and W. E. Diewert. 1982a. “Multilateral Comparisons of Output, Input, and Productivity Using Superlative Index Numbers.” Economic Journal 92(365): 73–86. Caves, D. W., L. R. Christensen, and W. E. Diewert. 1982b. “The Economic Theory of Index Numbers and the Measurement of Input, Output, and Productivity.” Econometrica 50(6): 1393–1414. Charnes, A., W. W. Cooper, and E. Rhodes. 1978. “Measuring the Efficiency of Decision Making Units.” European Journal of Operational Research 2(6): 429–444. Chung, Y. H., R. Färe, and S. Grosskopf. 1997. “Productivity and Undesirable Outputs: A Directional Distance Approach.” Journal of Environmental Management 51(3): 229–240. Cropper, M. L., and W. E. Oates. 1992. “Environmental Economics: A Survey.” Journal of Economic Literature 30(2): 675–740. d’Arge, R. C., and K. C. Kogiku. 1973. “Economic Growth and the Environment.” The Review of Economic Studies 40(1): 61–77. Dakpo, K. H., P. Jeanneaux, and L. Latruffe. 2016. “Modelling Pollution- Generating Technologies in Performance Benchmarking: Recent Developments, Limits and Future Prospects in the Nonparametric Framework.” European Journal of Operational Research 250(2): 347–359. Diewert, W. E. 1992. “Fisher Ideal Output, Input and Productivity Indexes Revisited.” Journal of Productivity Analysis 3(3): 211–248. Färe, R., S. Grosskopf, B. Lindgren, and P. Roos. 1992. “Productivity Changes in Swedish Pharmacies 1980–1989: A Nonparametric Malmquist Approach.” Journal of Productivity Analysis 3(1–2): 85–101.

Productivity Measurement and the Environment 323 Färe, R., S. Grosskopf, C. A. K. Lovell, and C. Pasurka. 1989. “Multilateral Productivity Comparisons When Some Outputs Are Undesirable: A Nonparametric Approach.” Review of Economics and Statistics 71(1): 90–98. Färe, R., S. Grosskopf, M. Norris, and Z. Zhang. 1994. “Productivity Growth, Technical Progress and Efficiency Change in Industrialized Countries.” American Economic Review 84(1): 66–83. Färe, R., S. Grosskopf, and C. Pasurka. 2013. “Joint Production of Good and Bad Outputs with a Network Application.” In Encyclopedia of Energy, Natural Resources and Environmental Economics, edited by J. Shogren, Vol. 2, 109–118. Amsterdam: Elsevier. http://dx.doi.org/ 10.1016/B978-0-12-375067-9.00134-0. Farrell, M. J. 1957. “The Measurement of Productive Efficiency of Production.” Journal of the Royal Statistical Society, Series A, 120(III): 253–281. Fisher, A. C., and F. M. Peterson. 1976. “The Environment in Economics: A Survey.” Journal of Economic Literature 14(1): 1–33. Frisch, R. 1935. “The Principle of Substitution: An Example of Its Application in the Chocolate Industry.” Nordisk Tidskrift for Teknisk Økonomi 1(1–2): 12–27. Frisch, R. 1965. Theory of Production. Dordrecht: D. Reidel. Førsund, F. R. 1985. “Input-Output Models, National Economic Models, and the Environment.” In Handbook of Natural Resource and Energy Economics, Vol. I, edited by A. V. Kneese and J. L. Sweeney. Amsterdam: Elsevier Science Publishers BV. Chapter 8: 325–341. Førsund, F. R. 1999. “On the Contribution of Ragnar Frisch to Production Theory.” Rivista Internazionale di Scienze Economiche e Commerciali (International Review of Economics and Business) 46(1): 1–34. Førsund, F. R. 2009. “Good Modelling of Bad Outputs: Pollution and Multiple-Output Production.” International Review of Environmental and Resource Economics 3(1): 1–38. Førsund, F. R. 2010. “Dynamic Efficiency Measurement.” Indian Economic Review 45(2): 123–157. Also published as Chapter 4 in Benchmarking for Performance Evaluation: A Frontier Production Approach, edited by S. C. Ray, S. C. Kumbhakar, and P. Dua (2015), 187–219. London: Springer. Førsund, F. R. 2018. “Multi-Equation Modelling of Desirable and Undesirable Outputs Satisfying the Material Balance.” Empirical Economics 54(1), 67–99. doi: 10.1007/s00181-016-1219-016. Førsund, F. R., and L. Hjalmarsson. 1974. “On the Measurement of Productive Efficiency.” Swedish Journal of Economics 76(2): 141–154. Førsund, F. R., and S. Strøm. 1976. “The Generation of Residual Flows in Norway: An Input- Output Approach.” Journal of Environmental Economics and Management 3(2): 129–141. Førsund, F. R., and S. Strøm. 1988. Environmental Economics and Management: Pollution and Natural Resources. London: Croom Helm. Hartwick, J. M. 1977. “Intergenerational Equity and the Investment of Rents from Exhaustible Resources.” American Economic Review 67 (5): 972–974. Hicks, J. R. 1963 [1932]. The Theory of Wages, 2nd edition. London: Macmillan. Jaffe, A. B., and R. N. Stavins. 1995. “Dynamic Incentives of Environmental Regulations: The Effects of Alternative Policy Instruments on Technology Diffusion.” Journal of Environmental Economics and Management 29(3): S-43–S-63. Jaffe, A. B., S. R. Peterson, and P. R. Portney. 1995. “Environmental Regulation and the Competitiveness of U.S. Manufacturing: What Does the Evidence Tell Us?” Journal of Economic Literature 33(1): 132–163. Keeler, E., M. Spence, and R. Zeckhauser. 1972. “The Optimal Control of Pollution.” Journal of Economic Theory 4(1): 19–34.

324 Finn R. Førsund Kneese, A. V. 1971. “Background for the Economic Analysis of Environmental Pollution.” Swedish Journal of Economics (Special issue on environmental economics) 73(1): 1–24. Kneese, A. V., R. U. Ayres, and R. C. d’Arge. 1970. Economics and the Environment: A Materials Balance Approach. Baltimore, MD: Johns Hopkins University Press. Lanoie, P., M. Patry, and R. Lajeunesse. 2008. “Environmental Regulation and Productivity: Testing the Porter Hypothesis.” Journal of Productivity Analysis 30(2): 121–128. Lanoie, P., J. Laurent-Lucchetti, N. Johnstone, and S. Ambec. 2011. “Environmental Policy, Innovation and Performance: New Insights on the Porter Hypothesis.” Journal of Economics and Management Strategy 20(3): 803–842. Leibenstein, H. 1966. “Allocative Efficiency vs. ‘X-Efficiency.’” American Economic Review 56(3): 392–415. Leontief, W. 1970. “Environmental Repercussions and the Economic Structure: An Input- Output Approach.” The Review of Economics and Statistics 52(3): 262–271. Leontief, W. 1974. “Environmental Repercussions and the Economic Structure: An Input- Output Approach: A Reply.” The Review of Economics and Statistics 56(1): 109–110. Leontief, W., and D. Ford. 1972. “Air Pollution and the Economic Structure: Empirical Results of Input–Output Computations.” In Input–Output Techniques, edited by A. Brody and A. Carter, 9–30. Amsterdam; London: North-Holland. Mäler, K.-G. 1991. “National Accounts and Environmental Resources.” Environmental and Resource Economics 1(1): 1–15. Martin, R. E. 1986. “Externality Regulation and the Monopoly Firm.” Journal of Public Economics 29(3): 347–362. Meade, J. E. 1952. “External Economies and Diseconomies in a Competitive Situation.” Economic Journal 62 (245): 54–67. Mishan, E. J. 1971. “The Postwar Literature on Externalities: An Interpretative Essay.” Journal of Economic Literature 9(1): 1–28. Murty, S. 2015. “On the Properties of an Emission-Generating Technology and Its Parametric Representation.” Economic Theory 60(2): 243–282. Murty, S., and R. R. Russell. 2018. “Modeling Emission-Generating Technologies: Reconciliation of Axiomatic and By- Production Approaches.” Empirical Economics 54(1), 7– 30. doi: 10.1007/s00181-016-1183-4. Murty, S., R. R. Russell, and S. B. Levkoff. 2012. “On Modelling Pollution-Generating Technologies.” Journal of Environmental Economics and Management 64(1): 117–135. Nijkamp, P., and J. Paelinck. 1973. “Some Models for the Economic Evaluation of the Environment.” Regional and Urban Economics 3(1): 33–62. Nordhaus, W. D., and E. C. Kokkelenberg (eds.). 1999. Nature’s Numbers: Expanding the National Economic Accounts to Include the Environment. National Research Council. Washington, DC: The National Academies Press. doi: 10.17226/6374. Palmer, K., Oates, W. E., and P. R. Portney. 1995. “Tightening Environmental Standards: The Benefit-Cost or the No-Cost Paradigm?” Journal of Economic Perspectives 9(4): 119–132. Perman, R., Y. Ma, M. Common, D. Maddison, and J. McGilvray. 2011. Natural Resource and Environmental Economics, 4th edition. Harlow, UK: Pearson Education. Pethig, R. 2006. “Non- Linear Production, Abatement, Pollution and Materials Balance Reconsidered.” Journal of Environmental Economics and Management 51(2): 185–204. Pigou, A. C. 1920. The Economics of Welfare. London: Macmillan. Pittman, R. W. 1983. “Mulitilateral Productivity Comparisons with Undesirable Outputs.” The Economic Journal 93 (372): 883–891.

Productivity Measurement and the Environment 325 Plourde, C. G. 1972. “A Model of Waste Accumulation and Disposal.” The Canadian Journal of Economics /Revue canadienne d’Economique 5 (1): 119–125. Porter, M. E. 1991. “America’s Green Strategy.” Scientific American 264(4), 168. Porter, M. E., and C. van der Linde. 1995. “Toward a New Conception of the Environment- Competitiveness Relationship.” Journal of Economic Perspectives 9(4): 97–118. Shephard, R. W. 1953. Cost and Production Functions. Princeton, NJ: Princeton University Press. Shephard, R. W. 1970. Theory of Cost and Production Functions. Princeton, NJ: Princeton University Press. Smith, V. L. 1972. “Dynamics of Waste Accumulation: Disposal versus Recycling.” The Quarterly Journal of Economics 86(4): 600–616. Solow, R. M. 1956. “A Contribution to the Theory of Economic Growth.” Quarterly Journal of Economics 70(1): 65–94. Solow, R. M. 1974. “Intergenerational Equity and Exhaustible Resources.” Review of Economic Studies: Symposium on the Economics of Exhaustible Resources 41(5): 29–46. Stern, N. H. 2006. The Economics of Climate Change. HM Treasury. http://hm-treasury.gov.uk. Stigler, G. J. 1976. “The Xistence of X-Efficiency.” American Economic Review 66(1): 213–216. Stiglitz, J. E., A. Sen, and J.-P. Fitoussi. 2009. Report by the Commission on the Measurement of Economic Performance and Social Progress. http://graphics8.nytimes.com/packages/pdf/ business/ Stiglitzreport.pdf United Nations (UN). 1993. System of Accounts 1993. United Nations Department for Economic and Social Information and Policy Analysis, Statistics Division. New York: United Nations. Weitzman, M. 1976. “On the Welfare Significance of National Product in a Dynamic Economy.” Quarterly Journal of Economics 90(1): 156–162. Xepapadeas, A., and A. deZeeuw. 1999. “Environmental Policy and Competitiveness: The Porter Hypothesis and the Composition of Capital.” Journal of Environmental Economics and Management 37(2): 165–182.

Pa rt I I I

M IC ROE C ON OM IC ST U DI E S

Chapter 9

Produ ctivi t y a nd F inancial Performa nc e Emili Grifell-Tatjé and C. A. Knox Lovell

9.1. Introduction In this chapter we explore the complex relationship between business productivity and financial performance, and we use BHP Billiton, an Anglo-Australian global resources company headquartered in Melbourne, Australia, to illustrate key concepts.1 The BHP Billiton 2016 Annual Report provides a good introduction to the material we cover in this chapter. The Report contains three key performance indicators that are used “to assess the financial performance of the Company . . . and to make decisions on the allocation of resources,” one being underlying EBIT (earnings before interest and taxes and excluding exceptional items).2 Figures 9.1 and 9.2 track recent trends in profit (underlying EBIT, expressed in USD) and return on assets (underlying EBIT/ total assets). Segments of the Report are devoted to a discussion of the sources of profit variation, both through time and across businesses. The three primary sources are volume changes, price changes, and change in external factors such as exchange rate movements and climate change. Volume changes are attributed toproductivity change and growth. Price changes are attributed to changing market conditions in countries where commodities are consumed and produced. Importantly, the financial contribution of productivity improvements is attributed to both “sustainable productivity-led volume improvements” and “sustainable productivity-led cost efficiencies.” These productivity gains are valued at more than USD 10 billion over 2013–2016, values we refer to as a productivity bonus throughout the chapter. The concepts of productivity and financial performance are linked throughout the Report. It is apparent that BHP Billiton management understands that change in financial performance is driven by quantity changes, price changes, and change in external factors, and that productivity change creates value through its impact on both quantity

330 Emili Grifell-Tatjé and C. A. Knox Lovell EBIT 35000 30000 25000 20000 15000 10000 5000 0

2006

2007

2008

2009

2010

2011

2012

2013

2014

2012

2013

2014

2015

2016

Figure 9.1. Profit at BHP Billiton.

EBIT/TA 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0

2006

2007

2008

2009

2010

2011

2015

2016

Figure 9.2. Return on assets at BHP Billiton.

changes and unit cost changes. In this chapter we analyze the separate impacts of quantity change and price change on three measures of financial performance. In section 9.2 we measure financial performance with profit, whatever its precise definition, although we use EBIT. We develop an analytical framework that allows us to attribute profit change to quantity changes, a driver of which is productivity change, and price changes, a driver of which is price-recovery change. We measure productivity change with the ratio of an output quantity index to an input quantity index, and we

Productivity and Financial Performance 331 measure price-recovery change with the ratio of an output price index to an input price index.3 The analytical framework has a rich history, the primary contributors being Davis (1955); Kendrick (1961, Chapter 5; 1984); Kendrick and Creamer (1961); Kendrick and Sato (1963); Vincent (1968); Courbis and Templé (1975); Houéry (1977); writers associated with the French state agency Centre d’Étude des Revenus et des Coûts (CERC, 1969); Eldor and Sudit (1981); and Miller (1984). We utilize two accounting conventions within this framework. In subsection 9.2.1 we treat profit as the difference between revenue and cost, which can be positive, zero, or negative. In subsection 9.2.2 we treat profit as a return to those who bear the risk of providing capital to the business; in the business accounts this return augments the cost of capital, and so revenue minus augmented cost is zero by construction.4 The latter treatment also is consistent with national accounts, in which receipts and expenditures balance; this accounting identity is exploited by Jorgenson and Griliches (1967). In section 9.3 we explore selected topics relevant to both accounting conventions, including the identification of potential drivers of productivity change, and the appropriation (or distribution) of the financial benefits of productivity change among those agents involved in its creation, among others. In section 9.4 we measure financial performance with return on assets (ROA), and we embed the ROA analysis within a duPont triangle framework, in which change in ROA is decomposed into the product of change in profit margin and change in asset turnover. ROA is a widely acknowledged key financial performance indicator, and Kline and Hessler (1952) describe the pioneering use of ROA and its two components at duPont. In his study of management accounting at duPont, Johnson (1975, 185) describes the triangle as the use of accounting data “for management control” and to support “the allocation of new investment among competing economic activities.” This assessment of ROA is strikingly similar to the assessment of EBIT at BHP Billiton. The literature linking ROA as a financial performance indicator with potential drivers of ROA is extremely diverse, although apart from the recent work of Bloom and his colleagues (e.g., Bloom and Van Reenen 2007, 2010; Bloom, Lemos, Sadun, Scur, and Van Reenen 2014), the list of potential drivers of ROA seems to have managed to exclude productivity! Our analytical framework incorporates productivity change as a driver of ROA change. Inspired by the significance that BHP Billiton attaches to productivity- led cost reductions, in section 9.5 we measure financial performance with unit cost if we can aggregate multiple outputs into a single value, or with unit costs of individual outputs if we cannot. BHP Billiton identifies five business segments, one for each commodity it extracts and markets. The largest segment is iron ore, for which “[o]ur focus remains on producing at the lowest possible cost. . . .”5 Nearly a century ago, Bliss (1923, 104) recommended the use of unit cost, particularly “[i]n businesses having satisfactory measures of physical volume.” Gold (1971) believed information on unit cost to be essential to all areas of managerial decision-making, including the allocation of resources among product lines and pricing policies. More recently, Borenstein and Farrell (2000) have explored alternative cost-cutting strategies. Each of these scenarios fits BHP Billiton.

332 Emili Grifell-Tatjé and C. A. Knox Lovell Section 9.6 provides a summary of the chapter and suggests some avenues for future research.6 We conclude this section with an introduction to our notation, which we augment in the following as necessary. A firm uses input vector x = ( x1 , …, x N ) ∈R+N to produce output vector y = ( y1 , …, y M ) ∈R+M . The set of feasible (x, y) combinations is the production set T = {(x, y): x can produce y}, whose outer boundary is a production frontier. The set of feasible x is the input set L ( y ) = x : ( x , y ) ∈T ∀y , and the set of feasible y is the output set P ( x ) = y : ( x , y ) ∈T ∀x . Input prices and output prices are given by the vectors w = (w1 , …, w N ) ∈R+N+ and p = ( p1 , …, pM ) ∈R+M+ . Cost is C = wT x = Σ nN=1wn xn ≥ 0, revenue is R = pT y = Σ mM=1 pm ym ≥ 0, profit is π = R − C > < 0, > and profitability (or cost recovery) is Π = R / C 1. Return on assets is π / A, where A < is the firm’s assets, and the duPont triangle decomposes ROA as π/A = π/R × R/A, the product of the profit margin and asset turnover. A cost frontier is defined as c y , w = min x wT x : Di y , x ≥ 1 ≤ wT x , in which the input distance function Di ( y , x ) is defined as Di ( y , x ) = max{φ : x / φ ∈L ( y )} ≥ 1. A revenue frontier is defined as r x , p = max y pT y : Do x , y ≤ 1 ≥ pT y , with output distance function Do (x , y ) defined as Do (x , y ) = min{θ : y / θ ∈P ( x )} ≤ 1.7 We consider two time periods, indicated by superscripts “0” and “1.” Thus profit change from base period to comparison period is π1 – π0.The superscripts also can refer to producing units, for example a benchmarking organization and a target organization, in which case “change” becomes “variance” or “deviation.” Miller (1984) developed a framework similar to ours to analyze yet another difference, that between actual and anticipated comparison-period profit, and Grifell-Tatjé and Lovell (2013) apply this framework to cost variance analysis.

{

(

{

)

(

)

{

}

) }

(

{

}

(

) }

9.2. Decomposing Profit Change Decomposing profit change requires a definition of profit, which varies across accounting conventions. We consider two conventions. One accounting convention defines π = R – C, without imposing equality between the value of output and the value of input, and without associating profit with any specific input. This convention is consistent with defining the cost of capital narrowly as depreciation expense and interpreting R – C as profit (EBIT), and also with defining the cost of capital more broadly as depreciation expense plus interest expense, and interpreting R – C as “pure” profit (EBT). The choice between reporting interest as an expense or as a component of profit can be important empirically, but it is irrelevant analytically. Under this accounting convention, we analyze variation in π = R – C.

Productivity and Financial Performance 333 Another accounting convention places special emphasis on the capital input. We write wNxN = wNK, K being the capital input and wN being the unit cost of capital. Davis (1955) associates wNK with depreciation expense, which is consistent with BHP Billiton’s use of EBIT as a measure of profit. We distinguish the cost of capital wNK from the (endogenous) return to capital, which we define as rK=π, r being the (endogenous) rate of return to capital. The return to capital must be sufficient to cover interest payments and taxes (hence EBIT), and also dividends and retained earnings to provide for future growth. Consistent with national income accounting conventions, this convention imposes equality between the value of output and the value of input by treating the return to capital as an expense, yielding augmented cost C = w1 x1 + + w N −1 x N −1 + (w N + r ) K = w T x. Under this accounting convention, we analyze variation in R − C ≡ 0.8 The two conventions use the same data to analyze variation in financial performance, but they organize the data in different ways, and they yield complementary insights. We analyze variation in R – C in subsection 9.2.1, and we analyze variation in R − C in subsection 9.2.2.

9.2.1. Change in R – C Under the first accounting convention, we examine change in R – C, one expression for which is

(

)

(

)

(

)

)

(

)

(

)

π1 − π 0 = p0T y1 − y 0 − w 0T x1 − x 0 + y1T p1 − p0 − x1T w1 − w 0 , (9.1) in which we use base-period prices to weight quantity changes in what we call the quantity effect, and comparison-period quantities to weight price changes in what we call the price effect. The quantity effect decomposes in two different ways:

(

p0T y1 − y 0 − w 0T x1 − x 0 = w 0T x1 (YL / X L ) − 1 + π 0 (YL − 1) (9.2) −1 = p0T y1 1 − (YL / X L ) + π 0 ( X L − 1) ,

in which YL = p0T y1 /p0T y 0 is a Laspeyres output-quantity index, X L = w 0T x1 /w 0T x 0 is a Laspeyres input-quantity index, and YL /X L 1 is a Laspeyres productivity index. The first term on the right side of each equality is a productivity effect, and the second is a growth (or contraction) effect. In the first equality the productivity effect scales the rate of productivity change [(YL /X L ) − 1] by deflated comparison-period cost w0Tx1 to generate a productivity bonus, which measures the contribution, positive or negative, of productivity change to profit change, and which amounted to USD 0.4 billion for BHP Billiton in 2016. The growth effect scales the output growth rate (YL − 1) by base period profit π0 to generate what we call a growth bonus, which measures the contribution, again positive or negative, of output change to profit change. The growth effect evaluates the business strategy of replication, often referred to as the “McDonalds approach”

334 Emili Grifell-Tatjé and C. A. Knox Lovell (Winter and Szulanski 2001). It is worth noting that the productivity effect can create (or destroy) value in the absence of output growth, and the growth effect can create (or destroy) value in the absence of productivity growth. Garcia-Castro, Ricart, Lieberman, and Balasubramanian emphasize this dual source of value creation in Chapter 10. The second equality is interpreted similarly, although the two effects in the second equality are not generally equal to their counterparts in the first equality. The price effect also decomposes in two different ways:

y1T ( p1 − p0 ) − x1T (w1 − w 0 ) = p0T y1 (PP / WP ) − 1 + π1 1 − WP−1 (9.3) −1 = w 0T x1 1 − ( PP / WP ) + π1 1 − PP−1 ,

in which PP = y1T p1 / y1T p0 is a Paasche output-price index, WP = x1T w1 / x1T w 0 is a Paasche input-price index, and PP / WP 1 is a Paasche price-recovery index, which measures the extent to which a producer’s output price changes compensate for its input price changes.9 The first term on the right side of each equality is a price-recovery effect, and the second is an inflation (or deflation) effect. In the first equality, the price-recovery effect scales the rate of price-recovery change [( PP / WP ) −1] by deflated comparison- period revenue p0T y1 to generate a price-recovery bonus, which measures the contribution, positive or negative, of price-recovery change to profit change. The price-recovery effect can be interpreted as a financial reflection of a firm’s market power. A notable application of the price-recovery effect occurs under incentive regulation, in which the regulator can constrain the ability of regulated firms to recover cost increases through price or revenue caps. Agrell and Bogetoft analyze incentive regulation in Chapter 16 of this Handbook. The inflation effect scales the input-price growth rate 1 − WP−1 by comparison-period profit π1 to generate an inflation bonus, which measures the contribution of input-price change to profit change. The second equality is interpreted similarly, and again the two effects in the second equality are not generally equal to their counterparts in the first equality. An empirically relevant interpretation of both inflation effects is that, assuming π1> 0, “a little inflation is good for business.” Conversely, the inflation effects also suggest that a little deflation, such as that recently threatening the European Union, is bad for business and has induced the European Central Bank to adopt policies designed to stimulate moderate inflation.10 Expressions (9.2) and (9.3) provide four distinct decompositions of profit change. Each identifies productivity change and price-recovery change as potential drivers of profit change, and the choice among them depends on the objective of the analysis. Identifying the drivers of productivity change, and those of price-recovery change, requires tools from economic theory. We provide an input-oriented identification of the drivers of productivity change in section 9.3. As a concluding observation, it is noteworthy that in its Annual Report BHP Billiton, although it does not explicitly follow our methodology, does decompose annual change in underlying EBIT into volume and price effects. It also decomposes the volume effect into a productivity effect and a growth effect, and it decomposes the price effect into change in sales prices and change in price-linked costs. It also reports a third component

Productivity and Financial Performance 335 of profit change, which includes exogenous factors such as exchange rate movements, exploration and business development, and asset sales.

9.2.2. Change in R − C Under the second accounting convention, we examine change in R − C . This change is zero by construction, but the zero change nonetheless decomposes into offsetting quantity and price effects as

(R

1

) (

)

(

)

(

)

(

)

− R0 − C 1 − C 0 = p0T y1 − y 0 − w 0T x1 − x 0 − r 0 K 01 − K 0 + y1T p1 − p0 − x1T w1 − w 0 − K 01 r 1 − r 0 (9.4) = p0T y1 − y 0 − w 0T x1 − x 0 + y1T p1 − p0 − x1T w 1 − w 0 ,

(

(

)

(

)

(

)

(

(

)

)

(

)

)

in which K 01 is comparison-period capital, valued at base-period prices. Capital is the only quantity variable measured in monetary units, and so the nominal change in capital from one period to the next combines the effects of price change with those of quantity change. We eliminate the effect of price change by using capital’s real comparison period value K 01.11 Expensing the return to capital in the base period makes p0T y 0 = w 0T x 0, which simplifies the quantity effect in expression (9.4) to

(

)

(

)

p0T y1 − y 0 − w 0T x1 − x 0 = p0T y1 − w 0T x1 = w 0T x1 YL / X L − 1 (9.5) −1 = p0T y1 1 − YL / X L , and so the quantity effect is a productivity effect in which X L = w 0T x1 w 0T x 0 is a Laspeyres input-quantity index with price weights w 0. Under this accounting convention, there is no growth effect, even if YL ≠ 1 or X L ≠ 1, because the associated value weight analogous to π0 in expression (9.2) is p0T y 0 − w 0T x 0, which is zero by construction. The two expressions for the productivity effect are equal, even if YL X L ≠ 1, because they both equal p0T y1 − w 0T x1 in the first row of expression (9.5). Expensing the return to capital in the comparison period makes p1T y1 = w 1T x1, which simplifies the price effect in expression (9.4) to

(

)

(

)

(

(

)

)

y1T p1 − p0 − x1T w 1 − w 0 = − y1T p0 + x1T w 0 − 1 (9.6) = p0T y1 PP / W P 0T 1 −1 , = w x 1 − PP / W P

(

(

)

)

336 Emili Grifell-Tatjé and C. A. Knox Lovell = x1T w 1 x1T w 0 is a Paasche and so the price effect is a price-recovery effect in which W P input price index. Under this accounting convention, there is no inflation effect, even ≠ 1, because the associated value weight analogous to π1 in expression if PP ≠ 1 or W P 1T 1 (9.3) is p y − w 1T x1, which is zero by construction. The two expressions for the price- ≠ 1, because they both equal − y1T p0 + x1T w 0 in recovery effect are equal, even if PP W P the first row of expression (9.6). Under this accounting convention, special interest attaches to the return to those who provide capital to the business; Davis (1955) called them “investors.” However, the return to capital is profit under a different name, since in the comparison period p1T y1 − w1T x1 − r 1K 01 = 0, p1T y1 − w1T x1 = r 1K 01 = π1, and in the base period p0T y 0 − w 0T x 0 − r 0 K 0 = 0, p0T y 0 − w 0T x 0 = r 0 K 0 = π 0. Consequently, r 1K 01 − r 0 K 0 = π1 − π 0. Thus the following expression for change in the return to capital also provides a new expression for profit change

(

)

(

)

(

)

r 1K 01 − r 0 K 0 = p0T y1 − y 0 − w 0T x1 − x 0 − r 0 K 01 − K 0 (9.7) + r 0 K 01 − K 0 + y1T p1 − p0 − x1T w1 − w 0 ,

(

(

)

(

)

(

)

)

and adding and subtracting r 0 K 01 − K 0 to the right side yields

(

)

r 1K 01 − r 0 K 0 = p0T ( y1 − y 0 ) − w 0T x1 − x 0 (9.8) + r 0 K 01 − K 0 + y1T p1 − p0 − x1T w1 − w 0 .

(

)

(

)

(

)

Thus, under this accounting convention, change in the return to those who provide capital to the business has three components: that portion of profit change attributable solely to productivity change (from expression (9.5)), a return to capital expansion effect r 0 K 01 − K 0 , and a price effect based on (p, w) rather than p, w , which appears in expression (9.3). The return to capital expansion effect was introduced by Eldor and Sudit (1981), and is analogous to the growth effects π 0 (YL − 1) and π 0 ( X L − 1) in expression (9.2), although it collects the impact of growth in a single input, that of capital expansion. We conclude by returning to Davis (1955), who showed that comparison-period profitability valued at base-period prices

(

)

(

)

1 = y1T p0 x1T w 0 Π 0 (9.9) = YL X L

is the Laspeyres productivity index appearing in expression (9.5), and that comparison- period profit valued at base-period prices

π 10 = p0T y1 − w 0T x1 1 − 1 (9.10) = w 0T x1 Π 0

(

)

Productivity and Financial Performance 337 converts the Laspeyres productivity index in expression (9.9) to a Laspeyres productivity bonus. Since expression (9.10) is another way of expressing the second equality in expression (9.5), it confirms that, when profit is expensed, the quantity effect is a productivity effect (which we also call a productivity bonus). The significance of 1 and π 1 , both of these two results is that the two financial performance indicators Π 0 0 which can be obtained from a company’s accounts, provide measures of productivity change and the productivity bonus, respectively. These two results do not hold unless profit is expensed.

9.3. Selected Topics We briefly consider some topics relevant to both accounting conventions, and we show how the treatment of each differs between the two conventions.

9.3.1. Drivers of Productivity Change We have attributed a portion of profit change to productivity change, but we have not explored the sources of productivity change. Doing so requires specification of an orientation, either input-conserving or output-expanding, which in turn depends on management strategy. We adopt an input-conserving orientation, in keeping with our observation in section 9.1 on the emphasis that BHP Billiton places on the cost savings arising from productivity growth. We develop two approaches. In the first approach we begin with the first productivity effect in expression (9.2), which we rewrite, exploiting the fact that Π 0 = p0T y 0 / w 0T x 0 , as

( = (p

) ( y − y ) − w (x /Π ) y −w x .

w 0T x1 (YL / X L ) − 1 = p0 / Π 0

0

0

T

T

1

1

0T

0

0T

1

1

− x0

) (9.11)

With the assistance of Figure 9.3, which depicts production sets in base and comparison periods, we decompose the productivity effect in expression (9.11) as

(p

0

Π0

)

T

y1 − w 0T x1

( (x

) )

(

= w 0T x 0 − x A − w 0T x1 − x C +w

0T

A

−x

B

(

)

technical efficiency effect technical change effect

)

+ ( p0 / Π 0 )T y1 − y 0 − w 0T (x C − x B ),

(

)

(9.12)

size effect

in which x A = x 0 Di0 y 0 , x 0 is a technically efficient radial contraction of observed base- period input vector x 0 , x B = x 0 Di1 y 0 , x 0 incorporates a further radial contraction of

(

)

338 Emili Grifell-Tatjé and C. A. Knox Lovell y

T1

y1

T0 y0

xB

xA

x0

xC

x

x1

Figure 9.3. Identifying the drivers of productivity change.

(

)

x0 made possible by input-saving technical progress, and x C = x1 Di1 y1 , x1 is a technically efficient radial contraction of observed comparison-period input vector x1. If production is more technically efficient in the comparison period than in the base period, the technical efficiency effect contributes positively to productivity change. If technical change is input-saving technical progress, the technical change effect also contributes positively to productivity change. Size change can contribute to productivity change in either direction, depending on the magnitudes of y1 − y 0 and x C − x B , and since y1 and xC are not necessarily radial expansions or contractions of y0 and xB, size change includes both scale and mix changes. Decomposition (9.12) has an input-saving, cost-reducing orientation, in the sense that both technical efficiency change and technical change are measured, and valued, in an input-saving direction. It is analytically possible, and equally plausible from a managerial perspective, to adopt an output-and revenue-enhancing orientation when decomposing the productivity effect, so that both technical efficiency change and technical change are measured and valued in an output-enhancing direction. In the second approach, originally developed by Grifell-Tatjé and Lovell (1999), we also begin with the first quantity effect in expression (9.2), but we use the entire quantity effect instead of just its productivity effect component. Continuing to use Figure 9.3, we have

(

(

)

(

p0T y1 − y 0 − w 0T x1 − x 0

=w

0T

+w

0T

+ p0T

(x (x (y

0

A 1

) − w (x −x ) − y ) − w (x −x

A

0T

)

1

B

0

0T

C

− xC

)

(

)

)

technical efficiiency effect

)

activity effect

− xB ,

technical change effect

(9.13)

Productivity and Financial Performance 339 in which the technical efficiency effect and the technical change effect are unchanged from expression (9.12). The activity effect collects the margin effect from the first quantity effect in expression (9.2) and the size effect in expression (9.12), and quantifies the aggregate impact on profit of efficient firm growth along the surface of T1 in Figure 9.3. In this approach, the technical efficiency effect and the technology effect are the only drivers of productivity change, which is defined as w 0T x 0 − x B − w 0T x1 − x C . This strategy of decoupling the impacts of productivity change and size change on profit change can be particularly appropriate in certain situations, such as

(

)

(

)

• when one of the main components of the business model and strategy of the firm is growth. In this case, it is appropriate to distinguish the contribution of growth from that of productivity change to change in financial performance. The activity effect plays a leading role in the Brea-Solís, Casadesus-Masanell, and Grifell-Tatjé (2015) study of Walmart’s sources of competitive advantage. • when a regulator is willing to pay for improvements in productivity associated with technical efficiency change and technical change, but not for size change linked with mergers and acquisitions. De Witte and Saal (2010) distinguished the activity effect from the productivity effect in their study of regulatory impacts on the Dutch drinking water sector. The decomposition strategy is very similar under the second accounting convention, in which R − C = 0. The productivity effect in expression (9.5) can be rewritten as

(

)

(

)

(

)

w 0T x1 YL / X L − 1 = p0T y1 − y 0 − w 0T x1 − x 0 , (9.14) and we apply the same decomposition as in expression (9.12), substituting w 0 for w0. The size effect in expression (9.12) coincides with the activity effect in expression (9.13) since p0 Π 0 = p0 / p0T y 0 / w 0T x 0 = p0. Consequently, under the second accounting convention, decompositions (9.12) and (9.13) coincide apart from their different input price vectors. This decomposition identifies the same input vectors xA, xB and xC, but it weights output-quantity changes with p0 rather than (p0/Π0) and weights input-